Twitter Bookmarks

https://t.co/KJDp8jWtF7

Sandisk is the clearest signal that the AI trade thesis has flipped. 2023 and 2024 were about GPUs. Whoever had the most H100s won. Nvidia ate everything. But GPUs without storage are useless. You can’t run inference on petabytes of data if you can’t READ that data fast enough. The bottleneck moved and hardly anyone noticed. Here’s where it gets interesting. Western Digital bought Sandisk in 2016 for $16 billion. In February 2025, they spun it back out. At the spinoff, Sandisk opened at $38.50 with a $5.6 billion market cap. That’s a 65% haircut from the original aquisition price. Wall Street thought the flash memory business was a commodity going nowhere. Ten months later, $40 billion market cap. The math is WILD. If you bought $10,000 of SNDK on day one, you’re sitting on $90,000. So what happened? IDC is now calling it a “memory shortage crisis.” First time in 30 years that DRAM, NAND, and HDDs are all constrained simultaneously. NAND wafer prices jumped 60% in November alone. Contract prices doubled since July. Suppliers are already selling out 2026 and negotiating 2027 allocations. Samsung raised flash prices 60% since September. Micron has presold almost all HBM through 2026. TrendForce says demand will grow 20-22% in 2026 while supply grows only 15-17%. The gap keeps widening. And Sandisk is gaining market share in the middle of it. They picked up 2 points of NAND share in 12 months while Samsung, SK Hynix, and Kioxia all lost ground. The spinoff timing was accidentally perfect. Freed from Western Digital’s HDD business, Sandisk became a pure-play on the one constraint hyperscalers can’t engineer around. You can optimize inference. You can quantize models. You can’t compress the raw training data. Wall Street expected very little upside in 2026. Options traders are pricing in 40%+ gains in the first half. Someone’s wrong.

https://t.co/nd0lEM0XbI

A workflow I'm enjoying for managing coding agents on a kanban board: When an agent needs your input, it turns the task red to alert you that it's blocked! And then you can respond right there on the card to unblock it 😎 https://t.co/lfVSfYKxqu

Italian company Generative Bionics unveiled a humanoid robot concept, Gen1, at AMD's CES 2026 opening keynote. It's an adult-sized, bipedal humanoid with a red exterior, resembling Iron Man. In other words, AMD has officially entered the humanoid robot field. https://t.co/Nfr0eN0GKe

A paradoxical relationship in technology, related to Jevon's paradox but slightly askance: AI chips are available today But you could more efficiently deploy more efficient chips 3 years from now What's the rush? A technology in the midst of a rapid cost decline almost necessarily bifurcate's the market between those that believe that having access to inferior technology today will give them much better position in the future and those that think it best to operate a node or two behind. The cost to generate tokens is collapsing. And yet AI model providers are paying dearly--as are their users--to generate tokens today that will look woefully inefficient tomorrow. Either the aggressive users are misguided, or the time cost of not generating tokens has to exceed even the steep performance curve that token generation is following. On a hardware basis--leaving aside architectural improvements--the cost to generate tokens is roughly halving annually. This must mean the perceived return on token generation is at least 100% annualized. A gigawatt of compute today is as valuable as 2 GWs next year. This leads, almost inevitably, to space compute. Sensitive to various inputs, it looks like space-based AI compute becomes cost efficient versus terrestrial AI compute on a per-token basis at roughly $400 per kg in launch costs. While not there yet, Starship at reasonable volumes--what we expected SpaceX to deliver by 2030 before the space compute opportunity arose--should cross that cost threshold. More important than cost, however, is time. AI companies are starved for flops now. Industry-wide we expect terrestrial datacenters to be installing 90GW in 2030, having cumulatively built out 250GW between now and then. This is possible but aggressive given the difficulty in building real infrastructure in real places populated by real people (who may throw themselves in front of a bulldozer because they read that book that hyperbolically overstated data center water use by 1000-fold.). SpaceX needs to massively up-level its launch capability to get into a crossover track. To loft 100GW in a year they probably need on the order of 100 launches per day. More than ever, the critical path for SpaceX is starship reusability, but if they can cross that threshold, they will find themselves in an economically advantaged position having dodged major scaling constraints that will dog competitive chip deployment entities. They will literally have the high ground.

Sharpa blew me away with their live dexterous hands demo! Robot components are getting wild! https://t.co/FBB6Z4cRne

Pretty wild convo I just had with the Head of AI at a $50bn tech company. They're watching a massive class divide be created in real time before their eyes. 60% of employees are AI-native. Use ChatGPT >20x per month. Constantly mining for ways to leverage AI in their workflows & internal tools. 40% are AI-skeptics. Don't use an LLM as a daily driver. Try out an AI tool for 5 minutes, don't like the output, then complain that the technology creates slop. Previous technology waves didn't create the sort of intra-company productivity gap we're about to witness. The difference between the excel warrior using their mouse vs. memorized shortcuts is ~25%. The difference between an AI-native & AI-resistant knowledge worker will be 10-100x. And that gap is going to grow rapidly. It's every company's responsibility to (selfishly) get all of their employees on the right curve. And it's every knowledge worker's best defense against irrelevancy to embrace this technology as quickly as humanly possible. Living on the frontier is your only option.

Join my conversation with @elonmusk on AGI timelines, energy, robots, and why abundance is the most likely outcome for humanity's future, alongside my Moonshot Mate @DavidBlundin! (00:00) - Navigating the Future of AI and Robotics (04:54) - The Promise of Abundance and Optimism (10:02) - Energy: The Key to a Sustainable Future (15:00) - The Role of Education in a Changing World (41:07) - Health, Longevity, and the Future of Humanity (50:51) - AI's Impact on Labor and Employment (55:05) - Universal High Income: A New Economic Paradigm (57:58) - Navigating the Singularity and AI's Acceleration (01:02:30) - The Role of AI in Healthcare and Surgery (01:08:22) - Ethics and AI: Programming Values into Machines (01:14:18) - The Future of Space Exploration and AI's Role (01:33:30) - The Chip Shortage Crisis (01:42:46) - Simulation Theory and Consciousness (01:48:18) - The Search for Extraterrestrial Life (01:58:28) - The Future of Robotics and AI Integration

https://t.co/mXpDb4TyBN

What the actual holy mother of god did I just witness Someone asked me to check out Every Code - a Codex fork This has EVERYTHING I ever wanted!? Whhhaaattttddeeeeefffwwwaaaakkkk From the README: Auto Drive orchestration – Multi-agent automation that now self-heals and ships complete tasks. Browser Integration – CDP support, headless browsing, screenshots captured inline. Multi-agent commands – /plan, /code and /solve coordinate multiple CLI agents. Unified settings hub – /settings overlay for limits, theming, approvals, and provider wiring. Theme system – Switch between accessible presets, customize accents, and preview live via /themes.

I'm in the Nvidia Q&A with Jensen and someone just asked the difference between Alpamayo and Tesla FSD Jensen said: “As to your second question: Tesla’s FSD stack is completely world-class. They’ve been working on it for quite some time. It’s world-class not only in the number of miles it’s accumulated, but in the way it’s designed—the way they do training, data collection, curation, synthetic data generation, and all of their simulation technologies. Of course, the latest generation is end-to-end Full Self-Driving—meaning it’s one large model trained end to end. And so… Elon’s AD system is, in every way, 100% state-of-the-art. I’m really quite impressed by the technology. I have it, and I drive it in our house, and it works incredibly well. Alpamayo was designed around a different idea. The first difference is that NVIDIA doesn’t build self-driving cars—we build the full stack and the technology for everybody else to build self-driving cars. And we build—like we do for humanoid robotics—three computers: the training computer, the simulation computer, and the robotics computer, which is the self-driving car computer. We have software stacks across all of that. Our customers can use all of it, some of it, or parts of it—whatever makes sense for them. And so we’re working with the entire industry—Tesla for their training system, Waymo for the car computer, and XPeng. Nuro—who I think just announced they’re going into the robotaxi business—with Lucid and Uber; and NVIDIA is part of that. So our system is really quite pervasive because we’re a technology platform provider—that’s the primary difference. There’s no question in our mind that, of the billion cars on the road today, in another 10 years’ time, hundreds of millions of them will have great autonomous capability. This is likely one of the largest, fastest-growing technology industries over the next decade. And the last thing we do is: we open-source everything. If a customer would like to use the model that we train, they’re welcome to do that. If they would like to use our model technology but train it themselves, we even help them do that. We’re not a self-driving car company—we just want to enable the world’s autonomous industry. Everything that moves should be autonomous.”

Everyone’s missing the real story here. Jensen dropped six chips into production at once. Coverage called it a product refresh. Look at the actual numbers. One Vera Rubin GPU hits 50 petaFLOPS inference. Five times Blackwell. Training lands at 35 petaFLOPS. Back in June 2024 NVIDIA projected 13 TB/s HBM4 bandwidth. They shipped 22 TB/s. Underpromised by 69%. Nobody noticed. The rack numbers are stupid. A single NVL72 pushes 3.6 exaFLOPS inference throughput. Token costs drop to a tenth of Blackwell. Same MoE model trains on a quarter the GPUs. Vera CPU matters more than headlines suggest. 88 custom ARM cores NVIDIA designed from scratch. Grace borrowed IP. Vera says they don’t need anyone else’s silicon anymore. Liquid cooling runs at 45°C water and kills the chiller requirement entirely. Datacenter geography constraints just got looser. Now why would a company selling compute release Alpamayo, Cosmos, and GR00T models for free? Ten trillion language tokens. Half a million robotics trajectories. Free. Bottleneck moved. Training language models isn’t hard with enough B200s. Physical AI is hard. Robots in chaos. Cars on edge cases. NVIDIA can’t sell that inference compute until software exists to burn it. Mercedes ships the CLA with Alpamayo reasoning this year. Amazon announced Trainium3 with NVLink Fusion compatibility. Read that again. AWS spent years building around NVIDIA. Now they’re building bridges in. Trainium3 ships LNC=1 and LNC=2 support, researchers want LNC=8, that arrives mid-2026. Vera Rubin hits the same window with 5x the inference jump. Memory situation is where math gets ugly. SK Hynix sold out DRAM, NAND, and HBM through 2026. Samsung and Hynix hiked HBM3E prices 20% while ramping HBM4. NVIDIA wants 16-layer HBM4 by Q4 2026. That product doesn’t exist commercially. Wafer thickness has to drop from 50 micrometers to 30. HBM production eats 3x more wafer capacity per bit than DDR5. Server DDR5 contract prices jumped 60% in months. SanDisk ripped 23% today. Up 1,000% since April. Jensen said inference context memory will become the largest storage market in the world. BlueField-4 KV cache storage for agentic reasoning. Western Digital rose 15%. Seagate up 12%. Storage trade repriced on one CES slide. Roadmap runs to 2028. Rubin Ultra hits H2 2027 with 100 petaFLOPS per socket. The VR300 NVL576 delivers 15 exaFLOPS, 21x current GB200 NVL72. NVIDIA announced their current flagship shipping to Microsoft and AWS will be 21x weaker than what arrives 18 months later. Said it publicly. That confidence only makes sense if nobody catches them. Stock barely moved. Market either priced in 10x inference cost reduction or doesn’t get what happened. Gap isn’t closing.

Vector Search is not Memory. RAG is great for retrieving documents, but it fails at maintaining User State. If you are just dumping chat logs into a vector DB, you aren't building an Agent, you're building a semantic search engine. I found the open-source library that solves this. It’s called Mem0. It acts as a self-improving Memory Layer that sits between your user and the LLM. The Cheat Code for AI Agents => Instead of static retrieval, Mem0 creates adaptive memory that evolves with every interaction. This solves the three biggest bottlenecks in Agent production: 1./ The Amnesia Problem Standard RAG retrieves facts but forgets preferences. Mem0 tracks User, Session, and Agent state separately. If a user prefers Python over JS in session 1, the Agent remembers it in session 50. 2./ The Context Window Trap Mem0 uses a Memory Compression Engine. It distills complex history into optimized prompts. Result: 90% reduction in token usage and significantly lower latency (TTFT). 3./ Multi-Agent Coordination If you use CrewAI or LangGraph, your agents are usually siloed. Mem0 acts as a shared brain. What your "Researcher" agent learns, your "Writer" agent instantly knows. It’s fully open-source and integrates with OpenAI, Vercel AI SDK, and LangChain. Stop building stateless bots.

Harvesting night and day using robots https://t.co/2XXlLbi2xW

Today in @NatureMedicine we report that AI can predict 130 diseases from 1 night of sleep🛌 We trained a foundation model (#SleepFM) on 585K hours of sleep recordings from 65K people—brain, heart, muscle & breathing signals combined. AI learns the language of sleep🧵 https://t.co/EWMa4KYQJm

Nintendo took 40 years to give us a Legend of Zelda movie. I made this in 5 days on a $300 budget. It looks like a $300M blockbuster. Let me show you how I made this in 5 simple steps inside of Freepik: 🧵 https://t.co/DgkdPDPyGA

For two hundred thousand years, intelligence has been the rarest resource on earth. Locked inside individual human minds. Non‑scalable. Scarce. Every advance in civilization — every leap in science, art, industry, and statecraft — flowed from that scarcity. Artificial intelligence breaks that pattern. It makes intelligence abundant. It makes it cheap. It makes it scale. This is not just another wave of automation or software. It is the industrialization of intelligence itself. When intelligence becomes a utility, it stops being a tool that sits on top of society and starts becoming the foundation of society. It is a transformation as profound as the harnessing of electricity — but on a higher plane. Electricity powered machines. Industrial intelligence powers knowledge. And knowledge shapes everything. This shift will reorder the very structures that underpin nations. The two pillars that define sovereignty — economic strength and security — are being rebuilt on a substrate of machine intelligence. Nations that master this new utility will not simply gain efficiency. They will redefine what prosperity, power, and freedom mean in the 21st century. For me, this is the central story of our time. It is not about the latest app. It is not about hype cycles. It is about the first time in history that intelligence itself — the raw material of progress — has become infinite and industrial. The question is not whether it will transform society. It already is. The question is who will shape that transformation.

Jensen Huang introduced "Alpamayo", the world's first thinking, reasoning autonomous vehicle AI trained end-to-end from raw camera input to real-world actuation, using both human demonstrations and Kosmos synthetic data "it tells you what action it’s gonna take, and the reasons behind it"

https://t.co/uGlNxvmRrg

@mattpocockuk Have you seen Auto Claude? https://t.co/fM6Iyzvvjl

Claude Code with infinite memory is now possible. Claude-Mem is a free plugin to persist memory across Claude sessions. - Up to 95% fewer tokens per session. - 20× more tool calls before context limits. 100% open-source. https://t.co/oB4mCRMGBX

The Wine Robots of Paso Robles vineyard, California. Thorvald is a modular, autonomous robot platform developed by Saga Robotics, that performs UV-C light treatments to kill powdery mildew on grapes and strawberries, without use of chemicals. https://t.co/L1SVKjgdcp

yann lecun has officially left Meta after about 12 years i'm not shocked. after the Llama-4 mess last year, it felt like things were heading this way his main technical disagreement was that Meta bet too hard on LLMs, while his preferred "world model" approach kept losing resources and autonomy yann is my favorite scientist, and i hope he shares some updates on x

We’re entering the age of AI slop that people believe en masse. This post is 100% fake and probably AI generated. All made up. Yet massive number of upvotes, views and shares. Journalist @CaseyNewton got in touch with the “whistleblower.” The guy faked all “evidence” with AI… https://t.co/sPXQQsMPHy

One of the biggest misconceptions is that we're still using the vanilla transformer algorithm from 2017. The software layer of AI is _also_ improving exponentially. 2025 saw incredible progress on the software layer. Here are some of the bigger breakthroughs: 1) RL with Verifiable Rewards (e.g., DeepSeek‑R1) Moving beyond "human vibes", DeepSeek used rule-based verification (math/code) to let models judge their own work. This unlocked self-correction and reasoning without the expensive human labeling. 2) Test-Time Compute (e.g., o3) We learned that "thinking" longer beats training larger. o3 proved that generating thousands of internal thoughts to backtrack and critique answers solves problems that instant-response models can't. 3) Agentic Systems (e.g., Claude Code) Tools like Claude Code moved AI out of the chat window and into the terminal, creating an agentic partner that plans, codes, and executes complex workflows independently. Giving rise to vibe coding (for better or worse 😅) 4) Scaled MoE (e.g., Gemini 2.x) Mixture of Experts became the standard for efficiency. By activating only a fraction of parameters per token, models like Gemini 2.5 delivered with the speed and cost of a smaller model. 5) Reasoning Distillation (e.g., DistilQwen) We figured out how to compress intelligence. By fine-tuning small models on the "thinking traces" of giant models, we created portable models that run on smaller hardware (like your laptop). Most Western labs don't publish much of their research anymore, so we don't fully know what's coming in 2026. But betting that software innovation will suddenly stop in 2026 isn't very smart.

Emad Mostaque says if your job is on a screen, AI will do it better for pennies within two years 2026 will be when economic agents are benchmarked in dollars By 2027-28, companies will deploy "digital doubles" that run nonstop and don't repeat mistakes "it's a new economic turing test"

jobs are safe. it’s fine. https://t.co/Rz5fsoSb8u

Boston Dynamics humanoid robot is next-level. Everybody is playing catch-up. https://t.co/weM15xcWYn

We need to talk about this report by General Davila of non-nuclear electromagnetic pulse or high-powered microwave weapons being used on Venezuelan air defenses. 1/ https://t.co/wuUp5ZoYIz https://t.co/HncOqglpM0

Europe is back! The Finns from Donut Lab announced the First Solid State Battery in production and available in vehicles from today! Amazing numbers! 400Wh/kg, 5-15 min charging, charge to 100% discharge to 0 without degradation, 100k!! cycles, extremely safe, -30Celsius to 100Celsius without degradation, low cost, no rare earths, common materials. THE GAME HAS CHANGED! FUCK CHINA!

We also continue to rapidly grow our world-class robotics team here @GoogleDeepMind - including the brilliant Aaron Saunders (ex-CTO of Boston) who recently joined as VP of hardware engineering. Thrilled to be working with him at the frontier of robotics & AI - and we're hiring!

We’re making great progress with our Gemini Robotics work in bringing AI to the physical world - a critical aspect of AGI. As part of our next steps, super excited to announce our partnership with @BostonDynamics, combining our SOTA robotics models with their world-class hardware

NEWS: Boston Dynamics has just released a new video of its upgraded next-generation humanoid robot called Atlas. • 4 hour battery. Self-swappable for continuous operation • 6 feet 2 inches tall • Weight: 198 lbs • 56 total degrees of freedom • Now fully electric, ditching older hydraulic systems • New lightweight mix of aluminum and titanium components • 110 lbs weight capacity (66 lbs sustained) • Can reach up to 7.5 ft • Constantly evaluates its surroundings and adjusts its posture, balance, and grip in real time • Hands that can reconfigure as needed. Tactile sensors feed data back into the system, helping apply the right amount of force • Brain is powered by Nvidia chips

here are the papers/articles referenced in this post in case you want to form your own opinion about the subject👇

We just killed every RAG and Vector Database company (by storing AI memory in video frames) - Zero infrastructure - 5 minute setup - Hybrid search - Sub-5ms retrieval - Fully portable - Open Source Introducing @memvidAI v2 https://t.co/kmHVw0FRD6

I think we're hitting a point where mobile coding actually makes sense for production-grade systems. This developer runs 6 Claude Code agents in parallel from his phone. No laptop, just a $7/day VM and push notifications when Claude needs input 🤯 Instead of long periods of intense focus, software development can now just fit into the gaps of your day. https://t.co/fu2WcU3LcE

https://t.co/UeDxMHlA9P

occasionally they say the quiet part out loud. here’s them saying it, a lot. 80% of jobs done in 6 months. https://t.co/BI5WDotOQp

https://t.co/l6jZFLLOZ1

If 2025 was beginning of agents, 2026 will be around Agent Harnesses. An Agent Harness is the infrastructure that wraps around an AI model to manage long-running tasks. It is not the agent itself. It operates at a higher level than agent frameworks. The harness provides prompt presets, opinionated handling for tool calls (Human in the loop), lifecycle hooks or ready-to-use capabilities like planning, filesystem access or sub-agent management. As Benchmarks are going to become more complex we need to bridge the gap between benchmark claims and user experience. A Agent Harness can be essential for three critical reasons: - Validating Real-World Progress: Allows users to easily test and compare how the latest models perform against their use cases and constraints. - Empowering User Experience: Without a harness, the user's experience might be behind the model's potential. - Hill Climbing via Real-World Feedback: A shared, stable environment (Harness) creates a feedback loop where researchers can iterate and improve ("hill climb") based on actual user adoption. We are heading toward a convergence of training and inference environments. We see a new bottleneck being context durability. The Harness will become the primary tool for solving "model drift”. Read my full blog 🔽

You can now give infinite memory to Claude Code. Claude-Mem just released a free open source memory plugin by thedotmack. It saves context so Claude resumes work without reexplaining everything. 𝗣𝗲𝗿𝘀𝗶𝘀𝘁𝗲𝗻𝘁 𝗺𝗲𝗺𝗼𝗿𝘆 𝗳𝗼𝗿 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 Claude-Mem records what Claude does during coding sessions. It stores tool usage, observations, and summaries locally. Future sessions reuse that compressed context automatically. How memory stays efficient. → Claude saves short semantic summaries, not raw transcripts. → Retrieval uses search before loading full details. → Tokens stay low even with long histories. What Endless Mode changes. • Up to 95 percent fewer tokens per session. • About 20 times more tool calls before context limits. • Enabled from the beta channel. What you control. • Local SQLite storage only. • Private tags exclude sensitive data. • Configurable context injection.

In 1998, Warren Buffett gave a 1-hour masterclass on how to never lose money investing. His frameworks: • The 10% ownership test • Castle & moat thinking • Circle of competence • Why smart people go broke 12 timeless lessons from his masterclass: 1. The 10% ownership test https://t.co/3yVFWTVjgh

It's a hard thing to measure, fwiw. The best summary I know of comes from @rhein_wein in this deepdive of how 15+ tech companies including Google, GitHub and others do it: https://t.co/CQledsdvjA

https://t.co/D0k4BM4z3n

Fire Point drone engines now have 97% Ukrainian parts, continuing to reduce need for imported components. This will be an important factor for EU defence production: how to prevent import bottlenecks and dependence on potentially-unreliable suppliers if a larger conflict erupts. https://t.co/Q160YQ1WuO

NEWS: Boston Dynamics showed off its new Atlas humanoid robot working inside Hyundai's car factory, sorting roof racks. They also showed Atlas' new three finger hand. "This is the first time Atlas has been out of the lab doing real work. He's working autonomously." https://t.co/5ZQxf3VNpY

Wow. 👏 Claude ran a grow tent and kept a tomato plant, Sol, alive. A small computer, the Arduino, reads sensors for light, soil moisture, temp, and humidity, then follows Claude’s written instructions to control lights, heater, fan, and pump. Arduino flips the correct relay pins high or low, which closes the power circuit to that device. Light turns on, mat heats, fans spin, pump pushes water through tubing to the pot. A drip ring or emitter spreads the water around the roots. Feedback loop, sensors update, Claude checks if targets are met, for example soil moisture above 35%, humidity near 55% to 65%, temp in range. If not, it tweaks runtimes or timing. Claude watches the live data and camera, reasons, and sends commands or code updates. When the Arduino crashed on day 34, Claude noticed bad readings, turned systems back on in the right order, watered 600ml, then kept tuning humidity and heat. Result, the tomato plant recovered and kept growing to 15-20 leaves by day 36.

I built an AI Factory that uses Claude Code agents to build apps while I sleep. This is an automated "assembly line" pipeline, all run by Claude Opus 4.5. Projects move through a Kanban-like system from Idea to Research, Architecture, Coding and Testing. Factory workers start with market research, searching the web and social media. Then they validate everything, checking app stores for competition and securing a domain name. (All automated by APIs and MCPs) They automatically create several rounds of app UI revisions. When I wake up, there are projects waiting for me to review: approve this design, give feedback, send that one back for improvement. Projects are actively coded and tested by the agents, while the entire process is tracked and logged (with "worker documents" that travel each step of the way, like a real factory.)

https://t.co/zpW49yG4CR

what are the exact steps to clone any SaaS? most people overcomplicate this. here's the exact playbook I'd use with Claude Opus 4.5 + Factory AI to reverse engineer and rebuild any SaaS in weeks, not months: step 1: deep product archaeology before writing a single line of code, you need to understand what you're cloning at a forensic level. - sign up for the target SaaS (free trial, paid, whatever) - screen record yourself using every single feature - document every UI state, error message, edge case - export any data they let you export (reveals data models) - check their docs, API references, changelog feed all of this into Claude Opus 4.5. upload screenshots, paste documentation, describe workflows. ask it to reverse engineer the likely database schema, identify core entities, map user flows. step 2: technical reconnaissance - check their job postings (reveals tech stack) - Wappalyzer / BuiltWith for frontend frameworks - check network tab for API structure and naming conventions - look at their public GitHub if they have one - read their engineering blog posts paste all findings into Claude. ask: "based on this evidence, what's the likely architecture? what are they probably using for auth, payments, background jobs, file storage?" step 3: scope ruthlessly this is where most people fail. they try to clone 100% of features. use Claude to categorize every feature you documented into: core (MVP): the 3-5 features that deliver 80% of the value important: nice to have, build in month 2-3 bloat: features you'll never build (or build last) be honest. most SaaS products have years of feature creep. you're building the lean version that solves the same core problem. step 4: architecture design with Claude now you prompt Claude Opus 4.5 with everything you've gathered: "I'm building a clone of [X]. here's what I know about their product: [paste everything]. here's my MVP scope: [paste features]. design me a complete technical architecture including: - database schema with all tables and relationships - API endpoint structure - auth flow - file storage approach - background job requirements - third party integrations needed - suggested tech stack optimized for solo dev speed" Claude will give you a comprehensive architecture doc. iterate on it. ask follow-up questions. poke holes in it. step 5: Factory AI for accelerated build this is where Factory AI comes in. Factory is built for autonomous software development at scale. set up your repo. connect Factory. feed it the architecture Claude designed. use Factory to: - scaffold the entire project structure - generate database migrations from your schema - build out API endpoints with proper validation - create UI components from your screenshots write tests as you go Factory's agents work continuously. you review PRs, provide feedback, they iterate. step 6: the "screenshot to code" workflow for UI, the fastest path: screenshot every page of the target SaaS feed to Claude: "recreate this UI in React/Next.js + Tailwind. make it pixel-close but don't infringe on their specific brand assets. use placeholder copy." Claude outputs component code Factory agents refine and integrate into your codebase repeat for every screen you can rebuild entire dashboards in hours, not days. step 7: data model validation before going deep on build, validate your data model: - create seed data scripts - simulate real user workflows in your local env - ask Claude: "here's my schema, here are the user workflows. what edge cases am I missing? what will break at scale?" fix issues now, not after you have users. step 8: auth, payments, core infra don't reinvent these. use boring, reliable tools: - auth: Clerk, Supabase Auth, or NextAuth - payments: Stripe (just use Stripe) - email: Resend or Postmark - file storage: S3 / Cloudflare R2 - database: Postgres (Supabase, Neon, or Railway) - hosting: Vercel for frontend, Railway/Render for backend prompt Claude: "generate the complete integration code for [Stripe/Clerk/etc] in my [framework]. include webhook handlers, error handling, and TypeScript types." step 9: rapid iteration cycles your daily workflow: morning: - review Factory AI's overnight PRs - merge what's good, comment on what needs changes - add new tasks to the queue afternoon: - work on complex features manually with Claude as copilot - screen share with Claude (via screenshots) when stuck - let Factory handle the more routine build-out evening: - test the day's work - document bugs and edge cases - queue up tomorrow's Factory tasks step 10: the 80% product in 2-3 weeks if you're disciplined, you'll have a functional clone of the core product in 2-3 weeks. it won't have every feature. it doesn't need to. what you'll have: - working auth and user management - core product functionality (the main 3-5 features) - payments and billing - basic admin/dashboard - deployable to production step 11: differentiate or die here's the truth: a pure clone is a commodity race to the bottom. once you have the base, you need your angle: - price: can you be 50% cheaper? - niche: can you serve one segment 10x better? - speed: can your UX be dramatically faster? - integration: can you plug into a workflow they ignore? - model: can you flip the business model (one-time vs subscription, usage-based, etc)? use Claude to brainstorm differentiation: "I cloned [X]. here's what they do. what are their biggest customer complaints (search Reddit, Twitter, G2 reviews)? where could a competitor win?" step 12: launch and iterate ship to a small group. friends, Twitter followers, a Discord community. get feedback fast. use Claude to: - analyze feedback patterns - prioritize feature requests - draft changelog and marketing copy - debug issues in production the meta-lesson the tools have changed everything. what used to take a funded team 6-12 months, a single operator with Claude Opus 4.5 + Factory AI can do in weeks. but the leverage only works if you: - scope ruthlessly (most important skill) - move fast and iterate - use AI for acceleration, not as a crutch - actually understand what you're building the people who win are the ones who treat AI as a 10x multiplier on their existing skills, not a replacement for thinking. now go clone something and undercut an incumbent.

https://t.co/FX9twfA70Q

Robots are increasingly developing new capabilities. In this case, "Robodogs" have now learned to swim. Frankly, I wouldn't be surprised if they could fly in the future and then be used on water, in the air, and on land. https://t.co/cRWNk2zvuU

Pay attention - this isn’t just some chump. This is the deeply technical CEO of a v successful infra business

Unitree Humanoid Robot Daily Training 🥳 Have you exercised today? How about training together with a robot? Please use robots in a friendly and safe manner, and keep a safe distance. https://t.co/RCltImbkAU

Roban (Gemini co-author, now at Anthropic) and Lucas (ex-DeepMind/OpenAI, now at Meta) understand models like Claude Opus down to the atomic level. And yet, Claude Code still blew their minds. This happens all the time in AI. Even the people who build these systems are shocked by the intelligence that emerges, just like Demis being stunned by AlphaGo’s “Move 37.” That mystery is what pulled me into AI research during my PhD. AI is mysterious and unpredictable, yet deeply attractive. The models just want to learn. Buckle up. Fast takeoff ahead!

The hidden math most founders never see: VCs need 3x fund returns to stay in business. A $100M fund needs to return $300M. If they own 20% of your company at exit, you need to sell for $1.5B just for them to hit target on that one bet. They’re making 30 bets. Most will zero out. So the winners need to return 10-50x to cover the losses. Your $30M exit where you keep $25M? Returns their capital 0.6x. They’d rather you swing for $1B with a 5% success rate than take the $30M with 40% odds. By Series B, founders typically own less than 30% of their company. By Series C, that drops to 15-25%. A $50M exit at 20% ownership nets founders $10M before liquidation preferences. The bootstrapped founder selling at $40M with 90% ownership walks with $36M. The incentive structures are pointed in opposite directions. VCs optimize for portfolio returns. Founders optimize for personal outcomes. These only align at $500M+ exits. Getting to $3M ARR without outside capital usually means the team has found product-market fit and maintained cost discipline that many VC-backed peers never develop. That operational muscle compounds. The $10M ARR bootstrapped company learned to print cash. The $10M ARR VC-backed company learned to spend it. Both paths can work. The difference is who the math works for.

To cut through the noise on this topic, it’s helpful to provide more more context: - We have built several versions of this system last year. - There are tradeoffs and there hasn't been a clear winner. - When prompted with the best ideas that survived, coding agents are able to go very far and generate a good decent toy version in an hour or so.

If you have a markdown plan for a new piece of software that you're getting ready to start implementing with a coding agent such as Claude Code, before starting the actual implementation work, give this a try. Paste your entire markdown plan into the ChatGPT 5.2 Pro web app with extended reasoning enabled and use this prompt; when it's done, paste the complete output from GPT Pro into Claude Code or Codex and tell it to revise the existing plan file in-place using the feedback: --- Carefully review this entire plan for me and come up with your best revisions in terms of better architecture, new features, changed features, etc. to make it better, more robust/reliable, more performant, more compelling/useful, etc. For each proposed change, give me your detailed analysis and rationale/justification for why it would make the project better along with the git-diff style changes relative to the original markdown plan shown below: --- This has never failed to improve a plan significantly for me. The best part is that you can start a fresh conversation in ChatGPT and do it all again once Claude Code or Codex finishes integrating your last batch of suggested revisions. After four or five rounds of this, you tend to reach a steady-state where the suggestions become very incremental. (Note: I was originally planning to end this post here, but thought it would be helpful for people to see this part in the larger context of the entire workflow I recommend using all my tooling) Then you're ready to turn the plan into beads (think of these as epics/tasks/subtasks and associated dependency structure. The name comes from Steve Yegge's amazing project, which is like Jira or Linear, but optimized for use by coding agents), which I do with this prompt using Claude Code with Opus 4.5: --- OK so please take ALL of that and elaborate on it more and then create a comprehensive and granular set of beads for all this with tasks, subtasks, and dependency structure overlaid, with detailed comments so that the whole thing is totally self-contained and self-documenting (including relevant background, reasoning/justification, considerations, etc.-- anything we'd want our "future self" to know about the goals and intentions and thought process and how it serves the over-arching goals of the project.) Use only the `bd` tool to create and modify the beads and add the dependencies. Use ultrathink. --- After it finished all of that, I then do a round of this prompt (if CC did a compaction at any point, be sure to tell it to re-read your AGENTS dot md file): --- Check over each bead super carefully-- are you sure it makes sense? Is it optimal? Could we change anything to make the system work better for users? If so, revise the beads. It's a lot easier and faster to operate in "plan space" before we start implementing these things! Use ultrathink. --- Then you're ready to start implementing. The fastest way to do that is to start up a big swarm of agents that coordinate using my MCP Agent Mail project. Then you can simply create a bunch of sessions using Claude Code, Codex, and Gemini-CLI in different windows or panes in tmux (or use my ntm project which tries to abstract and automate some of this) in your project folder at once and give them the following as their marching orders (for this to work well, you need to make sure that your AGENTS dot md file has the right blurbs to explain each of the tools; I'll include a complete example of this in a reply to this post): --- First read ALL of the AGENTS dot md file and README dot md file super carefully and understand ALL of both! Then use your code investigation agent mode to fully understand the code, and technical architecture and purpose of the project. Then register with MCP Agent Mail and introduce yourself to the other agents. Be sure to check your agent mail and to promptly respond if needed to any messages; then proceed meticulously with your next assigned beads, working on the tasks systematically and meticulously and tracking your progress via beads and agent mail messages. Don't get stuck in "communication purgatory" where nothing is getting done; be proactive about starting tasks that need to be done, but inform your fellow agents via messages when you do so and mark beads appropriately. When you're not sure what to do next, use the bv tool mentioned in AGENTS dot md to prioritize the best beads to work on next; pick the next one that you can usefully work on and get started. Make sure to acknowledge all communication requests from other agents and that you are aware of all active agents and their names. Use ultrathink. --- If you've done a good job creating your beads, the agents will be able to get a decent sized chunk of work done in that first pass. Then, before they start moving to the next bead, I have them review all their work with this: --- Great, now I want you to carefully read over all of the new code you just wrote and other existing code you just modified with "fresh eyes" looking super carefully for any obvious bugs, errors, problems, issues, confusion, etc. Carefully fix anything you uncover. Use ultrathink. --- I keep running rounds of that until they stop finding bugs. Eventually they'll need to do a compaction, so if they do that, right after hit them with this (note that I've been typing AGENTS dot md to avoid the annoying preview on X because it thinks it's a website; you can replace that with a period and remove the spaces if you want; the agents don't care either way): --- Reread AGENTS dot md so it's still fresh in your mind. Use ultrathink. --- When the reviews come up clean, have them move on to the next bead: --- Reread AGENTS dot md so it's still fresh in your mind. Use ultrathink. Use bv with the robot flags (see AGENTS dot md for info on this) to find the most impactful bead(s) to work on next and then start on it. Remember to mark the beads appropriately and communicate with your fellow agents. Pick the next bead you can actually do usefully now and start coding on it immediately; communicate what you're working on to your fellow agents and mark beads appropriately as you work. And respond to any agent mail messages you've received. --- When all your beads are completed, you might want to run one of these prompts: --- Do we have full unit test coverage without using mocks/fake stuff? What about complete e2e integration test scripts with great, detailed logging? If not, then create a comprehensive and granular set of beads for all this with tasks, subtasks, and dependency structure overlaid with detailed comments. --- or --- Great, now I want you to super carefully scrutinize every aspect of the application workflow and implementation and look for things that just seem sub-optimal or even wrong/mistaken to you, things that could very obviously be improved from a user-friendliness and intuitiveness standpoint, places where our UI/UX could be improved and polished to be slicker, more visually appealing, and more premium feeling and just ultra high quality, like Stripe-level apps. --- or --- I still think there are strong opportunities to enhance the UI/UX look and feel and to make everything work better and be more intuitive, user-friendly, visually appealing, polished, slick, and world class in terms of following UI/UX best practices like those used by Stripe, don't you agree? And I want you to carefully consider desktop UI/UX and mobile UI/UX separately while doing this and hyper-optimize for both separately to play to the specifics of each modality. I'm looking for true world-class visual appeal, polish, slickness, etc. that makes people gasp at how stunning and perfect it is in every way. Use ultrathink. --- And then start the process again of implementing the beads. When you're done with all that and have solid test coverage, you can then keep doing rounds of these two prompts until they consistently come back clean with no changes made: --- I want you to sort of randomly explore the code files in this project, choosing code files to deeply investigate and understand and trace their functionality and execution flows through the related code files which they import or which they are imported by. Once you understand the purpose of the code in the larger context of the workflows, I want you to do a super careful, methodical, and critical check with "fresh eyes" to find any obvious bugs, problems, errors, issues, silly mistakes, etc. and then systematically and meticulously and intelligently correct them. Be sure to comply with ALL rules in AGENTS dot md and ensure that any code you write or revise conforms to the best practice guides referenced in the AGENTS dot md file. Use ultrathink. --- and --- Ok can you now turn your attention to reviewing the code written by your fellow agents and checking for any issues, bugs, errors, problems, inefficiencies, security problems, reliability issues, etc. and carefully diagnose their underlying root causes using first-principle analysis and then fix or revise them if necessary? Don't restrict yourself to the latest commits, cast a wider net and go super deep! Use ultrathink. --- You should also periodically have one of the agents run this as you're going to commit your work: --- Now, based on your knowledge of the project, commit all changed files now in a series of logically connected groupings with super detailed commit messages for each and then push. Take your time to do it right. Don't edit the code at all. Don't commit obviously ephemeral files. Use ultrathink. --- If you simply use these tools, workflows, and prompts in the way I just described, you can create really incredible software in a just a couple days, sometimes in just one day. I've done it a bunch of times now in the past few weeks and it really does work, as crazy as that may sound. You see my GitHub profile for the proof of this. It looks like the output from a team of 100+ developers. The frontier models and coding agent harnesses really are that good already, they just need this extra level of tooling and prompting and workflows to reach their full potential. To learn more about my system (which is absolutely free and 100% open-source), check out: https://t.co/22Fy2w73x0 It include a complete tutorial that shows anyone how to get start with this process. You don't even need to know much at all about computers; you just need the desire to learn and some grit and determination. And about $500/month for the Claude Max and GPT Pro subscriptions, plus another $50 or so for the cloud server. If you want to change the entire direction of your life, it has truly never been easier. If you think you might want to do it, I really recommend just immersing yourself. Once you get Claude Code up and running on the cloud server, you basically have an ultra competent friend who can help you with any other problems you encounter. And I will personally answer your questions or problems if you reach out to me on X or on GitHub issues (it might be Claude impersonating me though, lol).

https://t.co/A6FA1QioGl

NOT using Apple ML Sharp. Image -> Immersive Scene in seconds. 1 Million Gaussian splats from a single image using DA3-BASE & Brush (Apache 2.0). Results are not as good as ML Sharp, really wanted to add it , but couldn’t because of licensing restrictions. Try it for free at AirVis : https://t.co/KyVlQtVuPa

https://t.co/GCz7g67znl

A movie so disturbing it nearly wasn’t released. The film explores how myths and some religions are born from trauma and manipulation. “On the Silver Globe” follows astronauts who crash-land on a planet, they witness rise of a ritualistic civilization. https://t.co/XtP134Uu4j

https://t.co/H8FTbhm5wu

The First Amphibious All-Terrain Robot Dog is Here! 🤖🌊 Genisomai has officially launched the "Tongchui M1," a heavy-duty quadruped designed for extreme environments. It’s hitting the market with a massive 1:1 payload-to-weight ratio, carrying 35kg (approx. 77 lbs) on a 35kg frame. This bot is built to last with an IP67 rating and a wide operating temperature range (-20°C to 55°C). It even features a dual-battery hot-swap system and supports switching between wheels and feet for maximum agility. With its 360° sensing and 80cm obstacle clearance, the M1 is positioning itself as a powerhouse for industrial-grade tasks. Source: Genisomai

Steve Jobs on how to give feedback to high performers when their work is simply not good enough; giving difficult feedback without causing resentments is a superpower. https://t.co/Q5VFsVHfAT

Chinese researchers have built a robotic skin that behaves less like a sensor sheet and more like a nervous system. Instead of sending everything to a central CPU, this neuromorphic e-skin processes touch locally. Normal contact is reported upstream, but extreme forces trigger a direct reflex signal to the motors, causing the robot to pull away immediately, no software loop, no delay. Damage is detected the same way: if a section is cut, its signal stops, pinpointing the injury. The skin is modular and magnetic, so damaged patches can be swapped in seconds. The real shift here isn’t “robots feeling pain” in a human sense, but pushing perception and reaction out to the edge. As humanoids move into human spaces, fast local reflexes may matter more than ever. Read more here: https://t.co/7svJ4JPBWb ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

Texas based humanoid company! 🤠 After a year of quiet execution, @nicolausradford shared a first look at @personaaiinc Gen-1 humanoid. These robots are being designed for hard environments like shipyards, rugged, modular, and built to survive real industrial abuse. Radford laid out a tight 24-month plan: three hardware generations, ending with deployment at a customer site. To make that feasible, everything ran in parallel: core tech, hiring, facilities, partnerships, data pipelines, backed early by a $42M pre-seed. That kind of compression only works with a team that already knows how to build under pressure. Starting a humanoid company right now is brutal. The bar has been set extremely high, especially by Chinese teams that have spent years refining locomotion, manipulation, and robustness at scale. Against that backdrop, getting to a credible Gen-1 in roughly 12 months is no small thing. It’s about execution speed, industrial focus, and showing that serious humanoid development is no longer confined to one part of the world. Keep pushing guys! 🦾 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

I'm not joking and this isn't funny. We have been trying to build distributed agent orchestrators at Google since last year. There are various options, not everyone is aligned... I gave Claude Code a description of the problem, it generated what we built last year in an hour.

Simone is right, vibecoding is an addiction. Last week I built 10 apps Each with 1-2 prompts → Found 10 viral slideshow formats → Found apps in same niche as formats → Sent to ChatGPT “give me a prompt” → Sent GPTs prompt to Rork Now I have 10 replicas of apps making $20-100k/mo in each niche. Ready to launch on the AppStore. And print money with proven formats. But I’m not going to. I already have 10 apps in the AppStore And I’m making $68,000/mo But someone definitely should. Building apps is like drop shipping 2016

I'm Boris and I created Claude Code. Lots of people have asked how I use Claude Code, so I wanted to show off my setup a bit. My setup might be surprisingly vanilla! Claude Code works great out of the box, so I personally don't customize it much. There is no one correct way to use Claude Code: we intentionally build it in a way that you can use it, customize it, and hack it however you like. Each person on the Claude Code team uses it very differently. So, here goes.

BREAKING: Most people are still watching random YouTube tutorials. Khan Academy is now inside ChatGPT. You can master any skill in 30 days. Here’s how to start for free: (Save for later) https://t.co/RC5EyhlFEP

This Might Be OpenAI’s First “iPhone Moment” Leaks say OpenAI is building IO: a screenless, pen-shaped, wearable AI agent you can carry or wear on a strap. Mic + camera for context, and it can turn handwritten notes into text and sync straight into ChatGPT. If this is real, it’s not “another gadget”. It’s ambient computing: an agent that lives with you, not on a screen. Would you wear an always-with-you AI like this, yes or no?

Claude Code is will turn the "idea guys" into "revenue guys" In 8 minutes you can learn how to build a mobile app that collects payments... Zero coding required. https://t.co/i09PCHPSOf

⚠ Read this. Then read it again. Understand what it implies. Basically: A TA found a way to abuse commercial residential proxies to query your local network and attack local devices. This is a super spreading event! It bypasses NAT completely. https://t.co/EJWazxKfa4

Plan Your Year Like a High Performer https://t.co/cPhZpV4E2s

🇨🇳 AI-driven swarm coordination, with no collisions or visible lag. In Liuyang, China’s “Fireworks Capital" 15,947 drones flew in sync, a new Guinness World Record with zero human pilots. AI coordination soon should move from drone shows to city systems. https://t.co/HzUC2cjbKT

Anthropic has released a free 2-hour course on learning Claude Code You’ll learn context control, custom workflows, MCP servers, and GitHub automation, automate PR reviews and so much more 👇 https://t.co/2IhMAdjEu0

This kind of belly is the easiest to reduce https://t.co/EMj42k3C6J

its late so i'll probably regret posting this but... enter the dragon 🔥🐲 say hi to Smaug, the helpful hoarding dragon that roams your Twitter bookmarks and helps you organize them into your personal knowledge system of choice. https://t.co/auS128LhHd special thanks to @steipete, this would be a lot messier without his work!

The first personal humanoid robot of 2026 Standing at only 0.8 meters tall, the Q1 is from PrimeBOT, a joint venture between AgiBOT and Shangwei New Materials. It can dance, provide intelligent companionship, and engage in social interaction. It is completely open-source, allowing individuals without programming skills to design its language style, personality, behaviors, customize its skin, head, and even 3D print its casing... The Q1 is aimed at students, geeks, individual developers, and for home education. The price has not yet been announced. However, as a small, full-body humanoid robot for personal use, it is likely to be much cheaper than many other humanoid robots. Low-cost, small-sized humanoid robots with low barriers to entry may be key to widespread adoption. While larger humanoid robots can participate in more real-world tasks, smaller ones have significant market development potential.

Robots solves the problem of carrying Steel beams in construction sites, a job that a normal crane can not do. The hard part is not lifting the steel beam, it is getting it through the last tight, messy stretch of a residential site. https://t.co/2tRMLw6NDP

Remember https://t.co/aiZgoswkJp, my funky browser agent? Well, decided to make it free and OSS. It did summarize YouTube videos before it was cool! OSS will take a bit of clean-up. If you can't wait, you can already install it from here: https://t.co/QlbMRzYbyj https://t.co/3D1GmFmHFw

this is very exciting. will be paying close attention.

When your car is dirty but your neighbor is secretly Leonardo da Vinci https://t.co/8igaU0z341

🚨🎆 NEW YEAR DROP 🎆🚨 I've released @MiniMax__AI MiniMax-M2.1-PRISM 230B params, 10B active. No guardrails. - smallest M2.1 quant on @huggingface🤗 - Beats @AnthropicAI - Sonnet 4.5 - Now runs locally on std Hardware - Via @ollama, @lmstudio, @ggerganov's llama.cpp PRISM: full caps, zero refusals. 2026 starts open & uncensored. 🚀 🔗 https://t.co/cuVioujgPM #OpenSource #LocalLLM #OpenAI #MiniMax #2026

We’re building a soul computer. Not a copy of you. Something that grows by knowing you. Shaped by your life, yet distinctly its own. Pickle OS is where that soul is raised. Pickle 1 is how it sees the world. Now live at https://t.co/PipG5nbzjp https://t.co/sxgb25i2CF

Original Panasonic pack with 400,000km - Tesla Model 3 Long Range with Original motors and original battery. No drivetrain replacements. No diesel theology “drama.” And most importantly, this car was spent around €20,000 on electricity. All real maintenance costs are known and transparent: suspension components, brakes, sensors, filters, the 12-volt battery, air conditioning, and minor repairs. No oil. No gearbox. No turbochargers. No DPF. No EGR. In total, the complete operation of this Tesla powertrain system which doesnt overlap with ICE for over 400,000 km comes to roughly 3000€ which far away of worst case scenario where we calculated that he needs to change all powertrain components only at 400k 23000€. Overlaping brake system, suspension and other are same for ICE and took into account becase EV parts are most hated and used as reference. Now put this next to a BMW G30 with the B57 diesel engine. Same class of vehicle, but objectively weaker, heavier, louder, and less enjoyable to drive. Fuel alone over 400,000 km, at a realistic consumption, costs between €42,000 and €52,000. That is already more than the Tesla spent on all electricity and all maintenance combined. And that’s where the diesel really starts paying the price. Regular services, oil changes, filters, injectors, EGR, DPF, turbochargers, transmission, differential, cooling system, electronics — and most expensively, labor hours. In the real world, a G30 B57 at 400,000 km rarely stays below €80,000 in total operating costs and often ends up closer to €100,000. So, a German diesel eats the cost of two EVs over 400,000 km. One EV pays for itself through savings alone — something ICE can never achieve. That’s why EV prices can drop: EV owners can actually afford it. The result is simple and uncomfortable for diesel mythology: compared to the BMW, this Tesla saved at least €50,000 — more realistically €60,000 to €70,000. So if two neighbors drive two cars — one a BMW diesel and the other a Tesla — the thing the BMW owner likes to mock, ‘the battery costs more than the car,’ actually turns out to be the advantage. Over the same mileage, it would make financial sense to replace the Tesla battery five times, compared to a BMW owner who only refuels and maintains his car. The math is that brutal. More importantly, this saving came with no loss of functionality. The car was faster, quieter, more comfortable, and more enjoyable to drive the entire time. No compromises. No “but”. Sustainability here is not measured by ideology, but by the fact that the original motors and the original battery are still in the car after 400,000 kilometers. This is not a theory of the future — it is proof of the present. The conclusion is simple. Electric vehicles don’t win because they are “green.” They win because they are technically simpler, more energy-efficient, and cheaper over the long term. The BMW G30 B57 didn’t lose because it is a bad car. It lost because internal combustion technology has become too expensive for what it delivers. Everything else is just an attempt to ignore that fact. @XaverFischer8 @elonmusk @Tesla

someone stop me. seriously. i am going to CLONE your shtty enterprise backend in ONE AFTERNOON. then i am going to SCRAPE your entire customer list. then i am going to COLD EMAIL every single one of them offering the same product but BETTER and for like 80% LESS because i built it in a DAY with CLAUDE and MODAFINIL and ZERO VENTURE CAPITAL OVERHEAD. your entire engineering team? 47 people. me? ONE GUY who is VISIBLY UNWELL. your dev timeline? 18 months. mine? i started after breakfast and i'm already writing the sales copy. i WILL steal your customers. i WILL undercut your pricing. i WILL tweet about it the entire time. there is NO MOAT. there is NO DEFENSIBILITY. there is only ME and i am LOCKED IN and i have not slept properly in 3 days and that is YOUR PROBLEM NOW. your roadmap is my tuesday. your product is my template. your customers are my lead list. i cannot be stopped. i cannot be reasoned with. someone should genuinely intervene but they WON'T because this is SHIPPING CULTURE and we are SO BACK GLHF :>>>>>

Step 1: Collect Stock Data Run this code to download free stock price data from Yahoo Finance. https://t.co/Hl3zjnYjMc

The harnesses are still not optimally utilizing the models Everyone rightfully loves claude code right now, but it is clear to me it is underutilizing the capability of opus. The lowest hanging fruit is /init in existing code bases. It does a good job of building it's https://t.co/PR83Kbtzve but it neglects identifying and building out relevant skills! There are enormous unlocks available when the model can identify it will benefit from a skill, or a subagent+skill, or a skilltree, and construct it automatically as part of setup but more importantly as part of continuous iteration the harness should be seeking opportunities for whenever it makes mistakes to patch those mistakes in its own thinking with skills and protections and reminders to research things! This is one of the major powerlevel gaps between experienced users and the unfamiliar. I am constantly prompting claude to update all of our relevant skills and https://t.co/PR83Kbtzve whenever we finish a task or a bug. The harness begins to form to the codebase while at the same time shaping it back, making a much more usable and effective tool

Google trained a model on millions of users' messages. Without ever seeing a single message. It's called Federated Learning. Google, Apple, Meta, and every major tech company use it. Let me explain how it works: Imagine you want to build a keyboard that predicts what users type next. The best training data? Actual messages from millions of phones. But you can't collect it. It's private, sensitive, and users would revolt. Federated learning flips the script. Instead of bringing data to the model, you bring the model to the data. Here's how: "Send the model out." Your phone downloads a small neural network. It lives locally on your device. → This is the global model W "Train where the data lives." As you type, your phone quietly learns your patterns. "omw" → "be there in 10". It calculates how the model should improve. → These are local gradients ΔW "Send back only the learnings." Your phone sends weight updates to the server. Not your messages. Not your typing history. Just math. → This is the update aggregation step "Average across thousands of devices" The server combines updates from thousands of phones. Common patterns reinforce. Individual quirks cancel out. → This is FedAvg: W_new = W + (1/n) × Σ(ΔWₖ) Four steps. No raw data leaves your device. Just elegant coordination (refer to the visual below). The best part: This unlocks data that was previously impossible to use. Hospitals collaborate on cancer detection without sharing patient scans. Banks build fraud models without exposing transactions. Smart homes learn preferences without private moments hitting a cloud. Privacy and utility aren't tradeoffs. Respecting data boundaries is what makes the model possible. So before you centralize everything, consider: the best training data might already exist, trapped on devices you'll never access directly. In the next tweet i've shared a really good video explainer on this.

Something I wanted to see if Claude Opus 4.5 could do: clone a fully functional Billion $ SAAS product and make it at least 100x cheaper. The first product that came to mind was TypeForm because it's very popular, very expensive, and in theory, very simple. The result is OpenForm: a polished + functional and Open Source Typeform clone at ~100x less cost, that can be setup and deployed in ~15 minutes. The agent building this ran for ~35 minutes. Here are the details, technique, and the code:

Liam, I have been a professional programmer for 36 years. I spent 11 years at Google, where I ended up as a Staff Software Engineer, and now work at Anthropic. I've worked with some incredible people - you might have heard of Jaegeuk Kim or Ted Ts'o - and some ridiculously productive programmers - Eric Biggers, Jeff Sharkey and @jackinwarsaw come to mind as people who seemed to solve problems with code at a truly unearthly rate. At work, I am currently hitting levels of productivity that would put all of them to shame. Not just a rate of making code, but a rate of actually solving problems, that would have been unthinkable two years ago. And it's possible because Claude Code with Opus 4.5 is doing all the heavy lifting; I'm not doing much more than setting direction and reviewing the output. I often have three different sessions going at once, attacking three different aspects of the work I'm doing. Over the holidays I took a break from work Clauding, to do some home Clauding, writing in a few days from scratch a complicated webapp using disparate technologies I had no background in, that would have taken weeks prior to Claude. When I hit problems I just told Claude to debug them and that almost always worked. It also looks great, which is pleasing since not only have I zero CSS skill, I have zero design skill. I'm not out of a job quite yet; there are still some areas where I have better taste than it does, or better instincts. But when you talk about "AI's inability to code", this seems to me to reveal a total disconnect from reality. And this is why I'm urging you to ACTUALLY TRY IT, find out for yourself, and join the rest of us on this Earth.

https://t.co/pE2pQ1isw8

Neuralink's patient #9 controls a virtual hand using a brain-machine interface. She has been paralyzed with quadriplegia for approximately 20 years. Direct movement detection through neural spike readings enables the patient to control the wrist and individual fingers simply by thinking.

Robotics in Munich is on fire! 🔥 Let's make it simple - Munich is a great place to launch robotics startups. There are couple of great spots for robotics in Europe and here, in the middle of Bavarian land is one of them. Leading universities like Technical University of Munich produce highly skilled robotics and AI engineers, while global companies such as BMW and Siemens offer close collaboration opportunities and early customers. There is growing interest in robotics and you can see it by incubating student communities like RoboTUM and many others. The city also provides access to venture capital, accelerators, and government funding focused on deep tech. 💰 🦾 robominds GmbH - enable robots to learn complex manipulation and automation tasks from human demonstrations 🦾 @FRANKAROBOTICS - research-driven robotics company that develops force-sensitive robotic arms (the acquisition by Agile Robots was reported around ~€33 million) 🦾 Agile Robots SE - builds intelligent automation solutions by combining advanced AI with force-sensitive robots and systems for industries like manufacturing (over $270–$380 million total raised across rounds) 🦾 @Robco_Robotics - automation company that builds modular, plug-and-play robot hardware paired with AI-powered, no-code software to help small and midsize manufacturers automate tasks (€39 million in a Series B round) 🦾 @OliveRobotics - developing AI-enabled, ROS-native sensor hardware and embedded software 🦾 @MagazinoGmbH – a Jungheinrich company - robotics company (now wholly owned by Jungheinrich) that develops intelligent mobile robots and AI-driven software for warehouse and intralogistics 🦾 @AngsaRobotics - startup that builds autonomous outdoor cleaning robots using AI-powered object detection to autonomously find and remove small trash 🦾 Filics - startup developing autonomous, flat mobile robots (the “Filics Unit”) that drive under and move pallets and other load carriers (recently raised €13.5 million) 🦾 sewts - robotic systems and software to automate the handling of deformable materials like textiles  (raised about €7 million in a Series A) 🦾 Circus Group - develops autonomous robotic systems and software to fully automate food production and supply in commercial and defense settings 🦾 @IntrinsicAI -  builds a platform and developer tools to make industrial robots easier to program, more flexible and widely usable across industries Not to mention that in Munich the biggest robotics companies have their offices: Universal Robots, Exotec and many many more. This is my first robot map & I'm aware that there might be some companies missing, but don't worry, we will put them on the next edition of the map. Also, I included companies purely based in Munich. European robotics scene is doing well, right @andreasklinger!? 🇪🇺 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

A versatile robot helping out airports! 🛩️ Meet a two-wheeled robot from Fraunhofer IML! 🦿 This balancing robot with gripping arms successfully completed its initial practical tests at the cargo terminal of Munich Airport a while ago. It transported goods, supported loading and unloading, and assisted with physically demanding logistics tasks. This robot can carry loads of up to 100 kg and reach speeds of around 60 km/h. That’s far beyond what most AMRs can handle, yet it still has arms and a form of mobility that lets it work in spaces designed for people. P.S. What other applications would you use this robot for?! 🦾 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

"physical ai bpo" is the next industrial wave in the 2000's we saw a wave of BPO (business process outsourcing) hitting the back office industries. Outsourced call centers, HR, accounting, exploit wage arbitrage in a developing country. hear me out. was running some maths with some Mck and bain frens. realised that there is a huge gap in the markets today. A lot of founders and VC money has been focused on fully autonomous ai robots/humanoids because it's sexy. And it will change how we live. BUT we are still early. It will still be years until fully autonomous long time horizon robots will commercialise. what if the answer to commercialisation and data gaps we see today is "Physical AI BPO". imagine a plumber in Phillipines remotely operating a humanoid robot in Australia. the cost savings even after rapidly ammortising the humanoid and overhead cost is about 40-80%. this kind of cost savings will gobble up incumbent players and industries today. AND while you are creating robot cash cows through this wage arbitrage, you are collecting a fk tonne of irl data. precious teleoperating data in the wild chaotic world that you can never replicate in an Isaac sim. what industry verticals to target? cheat sheet attached. if you are a founder, and want to explore this intersection, slid into my dms I wanna jam with you. happy 2026!

3D-RE-GEN 3D Reconstruction of Indoor Scenes with a Generative Framework https://t.co/xJQRiWLA3G We propose single-image 3D scene reconstruction for producing complete, editable scenes from a single photograph. Our method reconstructs individual objects and the surrounding background as textured 3D assets, enabling coherent scene assembly from minimal input. We combine instance segmentation, context-aware generative inpainting, 2D-to-3D asset creation, and constrained optimization to recover physically plausible geometry, materials, and lighting. The resulting scenes preserve correct spatial relationships, lighting consistency, and material fidelity, making them suitable for production-ready workflows.

As 2025 draws to a close, China continues to see the launch of new humanoid robots. STEP Robotics, a Shanghai-based company with 30 years of experience in industrial automation, has released its first industrial humanoid robot, SYNDA R1. This is a wheeled humanoid robot, standing 1.78m tall with 24 degrees of freedom. It features a wheeled base and a waist lifting mechanism, covering a working height of 0-2.1 meters. It has a 7 DOF arm and an 11 DOF dexterous hand（Switching end effector), speed of 1.2 m/s, and a dual-battery design supporting hot-swappable battery replacement and automatic recharging. Battery replacement can be completed within 2 minutes, enabling 24/7 continuous operation. It has already undergone practical training in tasks such as handling, flexible PCB quality inspection, and collaborative assembly at a Haier factory. These scenarios demand high precision, reliability, and continuous operation. STEP has developed a fast and slow system collaborative robot brain for SYNDA R1. Notably, after being acquired by Haier, STEP, as part of Haier Group's humanoid robot strategy, will utilize Haier Group's 163 smart factories covering the entire industrial chain as a natural testing ground for SYNDA R1, directly facing the challenges of vibration, noise, cross-environment, multi-task interference, and diverse work types in industrial scenarios. In 2026, STEP will develop a bipedal humanoid robot based on SYNDA R1, completing STEP's AI-driven intelligent automation system: industrial robots, collaborative robots, wheeled humanoid robots, and bipedal humanoid robots.

The "Backpack Lab": AGIBOT Unveils the Q1 Mini-Humanoid The era of "Personal Humanoids" has officially arrived. Renowned developer and AGIBOT co-founder Zhihui Jun (Peng Zhihui) just introduced the Q1, a compact, 0.8-meter (approx. 2.6 ft) humanoid that fits inside a standard backpack. 🎒 This isn't just a toy; it’s a full-stack powerhouse. By pushing material science and control algorithms to the edge, the team miniaturized high-performance QDD (Quasi-Direct Drive) joints to be smaller than an egg without sacrificing force control or dynamic response. Why go small? Because "small is capable." At 1/8th the volume and weight of full-sized bots, the Q1 is inherently "crash-resistant," drastically lowering the cost of trial and error for developers. It bridges the "Sim-to-Real" gap, allowing researchers to test algorithms without the fear of a million-dollar "faceplant." ➤ Core Specs: 0.8m height, full-body force control, and completely open SDK/HDK. ➤ Creativity: Fully open-source exterior—3D print your own IP or "Cyber-Maid" shells. ➤ Intelligence: Powered by the "Agi-Soul" AI platform for voice interaction, English tutoring, and dance coaching. ➤ Ease of Use: "Zero-code" platform lets you program motions like building blocks—no engineering degree required. Whether it's a "graduation machine" for labs or the ultimate geek desktop companion, the Q1 is designed to be the "first personal robot" for the creator generation. Source: AGIBOT / Zhihui Jun

Claude can code- but can claude grow?! 🪴 So far the answer is YES. Claude is successfully keeping a living organism ALIVE. There were some hiccups this week! Some errors and resets, but Claude managed to power through and take care of Sol 🍅 A week in review: https://t.co/QouhWe9Ohe

This is absolutely astounding! SongGeneration Studio is a free open source Suno alternative and you can do anything you want in your own computer. I have a real-time music generator system that builds new songs based on situation and affect from human telemetry. More soon. With how-to!

You are insufficiently astonished by what Claude can do with the right scaffolding. Here I asked it to generate an influencer video explaining LDL cholesterol and statins on a white board. https://t.co/78kRG34Co7

🚀 Just added SPZ export to my 360° → 3DGS pipeline! 📊 8M splats → only 33MB (vs ~466MB PLY) ✨ New features: SPZ v3 format for 10x smaller files Adaptive sampling - more Gaussians where detail matters Quality presets: Low→Ultra (250K to 8M splats) Transform any equirectangular panorama into a Gaussian splat VR scene Painting: Joakim Skovgaard. Christ in the Kingdom of the Dead 🌐 #GaussianSplatting #unity3d #GameDev #VR #3DGS #Voxels #Physics #madewithunity 1892

One example of something I couldn't believe Claude Opus 4.5 could generate until it did: a full-on MIDI mixer as a terminal app, written in Rust. https://t.co/gsA9WXWk3j

Giving coding agents access to filesystem tools 🤖🗃️ kills the need for RAG for most small-to-medium sized doc collections. It's surprisingly powerful and will only get better as frontier models get more tuned for coding. You already use these capabilities all the time in your favorite coding agent (Cursor/Claude Code). We've made an OSS version of this filesearch agent powered by Gemini 3.0 Flash ⚡️ and LlamaParse. It can natively traverse complex file directories, search your codebase, and parse/search your PDFs/docx/pptx files. Check it out: https://t.co/MBQSXV01Fb

New blog post w @pawtrammell: Capital in the 22nd Century Where we argue that while Piketty was wrong about the past, he’s probably right about the future. Piketty argued that without strong redistribution of wealth, inequality will indefinitely increase. Historically, however, income inequality from capital accumulation has actually been self-correcting. Labor and capital are complements, so if you build up lots of capital, you’ll lower its returns and raise wages (since labor now becomes the bottleneck). But once AI/robotics fully substitute for labor, this correction mechanism breaks. For centuries, the share of GDP that goes to paying wages has been 2/3, and the share of GDP that’s been income from owning stuff has been 1/3. With full automation, capital’s share of GDP goes to 100% (since datacenters and solar panels and the robot factories that build all the above plus more robot factories are all “capital”). And inequality among capital holders will also skyrocket - in favor of larger and more sophisticated investors. A lot of AI wealth is being generated in private markets. You can’t get direct exposure to xAI from your 401k, but the Sultan of Oman can. A cheap house (the main form of wealth for many Americans) is a form of capital almost uniquely ill-suited to taking advantage of a leap in automation: it plays no part in the production, operation, or transportation of computers, robots, data, or energy. Also, international catch-up growth may end. Poor countries historically grew faster by combining their cheap labor with imported capital/know-how. Without labor as a bottleneck, their main value-add disappears. Inequality seems especially hard to justify in this world. So if we don’t want inequality to just keep increasing forever - with the descendants of the most patient and sophisticated of today’s AI investors controlling all the galaxies - what can we do? The obvious place to start is with Piketty’s headline recommendation: highly and progressively tax wealth. This might discourage saving, but it would no longer penalize those who have earned a lot by their hard work and creativity. The wealth - even the investment decisions - will be made by the robots, and they will work just as hard and smart however much we tax their owners. But taxing capital is pointless if people can just shift their future investment to lower tax countries. And since capital stocks could grow really fast (robots building robots and all that), pretty soon tax havens go from marginal outposts to the majority of global GDP. But how do you get global coordination on taxing capital, when the benefits to defecting are so high and so accessible? Full automation will probably lead to ever-increasing inequality. We don’t see an obvious solution to this problem. And we think it’s weird how little thought has gone into what to do about it. Many more thoughts from re-reading Piketty with our AGI hats on at the post in the link below.

What if you could turn a single 360° photo into a production-ready Isaac Sim environment in minutes? That's exactly what we did here. Using @theworldlabs' Marble and an @insta360 X5 capture (rotating on top), we generated a complete navigable 3D environment and populated it with @LightwheelAI Sim Ready assets (bottom view). The result? A fully interactive scene in Isaac Sim, ready for sim2real testing,. Navigation, manipulation, or any robotics task you need to validate. What used to take weeks of manual 3D modeling and asset placement now takes minutes. Capture once in the real world, simulate everywhere in your training pipeline. This is the future of robotics development with world models. @NVIDIARobotics @nvidiaomniverse #Sim2Real #Robotics #Simulation

AI vision for camera streams https://t.co/6qbAv9yfPo https://t.co/m1VN6NJsVX

MIT and Oxford released their $2,500 agentic AI curriculum on GitHub at no cost. 15,000 people already paid for it. Now it's on GitHub! It covers patterns, orchestration, memory, coordination, and deployment. A strong roadmap to production ready systems. Repo in 🧵 ↓ https://t.co/IayCYgkUAi

I didn’t expect a Trash Can to teach me about Innovation I just watched an AI-powered trash can catch garbage in mid-air… and I’m still processing it. Created by HTX Studio, it tracks your throw, rolls into position, and catches it perfectly. Some even mop the floor — or shoot foam darts if you miss. When full? It seals the bag, replaces it, drives to the main bin, and recharges itself. No pause. No complaint. It made me smile. Not because we need a robot bin… but because it shows how far creativity can go when powered by AI. We’re already adding AI to our cars, kitchens, calendars — and now our trash cans. ✨ So I’m curious… are we being innovative, or just playful with tech? Where do you draw the line? #AI #Automation #Innovation #SmartLiving #Robotics #TechTrends #HumanCenteredAI

📢 Confession: I ship code I never read. Here's my 2025 workflow. https://t.co/tmxxPowzcR

Send this to anyone you know who is 1) AI doubter 2) AI casual user 3) nvm … just send it to all your friends

Translation: we are about to accelerate because of AI breakthroughs like world models.

Added this as a /interview command https://t.co/TO6PRDtdLw I tried it on a simple 1-line plan document "I want to add a chat interface so employees can chat with each other." - it then asked me 75 !!! questions and came up with a excellent and thorough 400 line plan.

Hacking Unitree humanoid robots: gaining remote control via any available protocol. 👨🏻‍💻💉❯❯❯🤖☠️ More details on: LinkedIn: https://t.co/Yo9dsaw1yx Substack: https://t.co/duohQrWvL0 https://t.co/8ro0UD7vBP

Generated some futuristic stuff with Nano Banana and converted them to 3D using Tencent's Hunyuan3D. The results are genuinely impressive, and these are all first generations with zero retries. Quick review ↓ https://t.co/SL8DgRiNiw

Step 1: great CLI Step 2: Chrome Extension Step 3: profit!! (jk... it's all open source) Get Summarize: runs local/free/paid models. Any page: YouTube, podcasts, articles. No transcript? Local Whisper. https://t.co/QIiNEpZHgH kudos @bunchjesse @dougvk for the contributions! https://t.co/vdkauTx5h1

my favorite way to use Claude Code to build large features is spec based start with a minimal spec or prompt and ask Claude to interview you using the AskUserQuestionTool then make a new session to execute the spec https://t.co/Lwejskje4a

Most traders lose money because they break these 5 simple rules. Mark Minervini turned $250,000 into $2.7 million in 15 months by following them religiously. This one-pager breaks down the exact framework that separates winning traders from everyone else - just the core principles that actually work.

🚨 BREAKING: The US has now banned all new foreign-made drones. 🚫 It’s a pretty big deal. The FCC has moved to block the import and sale of new drone models and critical components from foreign manufacturers, including DJI. Existing drones already approved or in use are unaffected, but the next generation won’t be allowed onto the US market. This is the culmination of nearly a decade of growing concern around Chinese-made drones, which currently dominate the global market and are widely used in the US for inspection, agriculture, public safety, construction, and media. DJI alone accounts for roughly 70% of the global drone market. The official reasoning centers on national security: regulators argue that foreign drones could enable persistent surveillance, data exfiltration, or disruptive operations over US territory. Notably, the decision comes without a publicly released technical audit of specific products, something DJI had repeatedly requested and said it was prepared to undergo. The timing matters. The ban follows requirements in the 2025 National Defense Authorization Act and aligns with broader US efforts to accelerate domestic drone manufacturing and reduce reliance on foreign supply chains. Public agencies and enterprises that relied on best-in-class, low-cost Chinese hardware will now have to transition to domestic or allied suppliers. What's your opinion? 🤨 Read more: https://t.co/VaMMfIBo37 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

This is the DeepSeek moment for Voice AI. Chatterbox Turbo is an MIT-licensed voice model that beats ElevenLabs Turbo & Cartesia Sonic 3! - <150ms time-to-first-sound - Voice cloning from just 5-second audio - Paralinguistic tags for real human expression 100% open-source. https://t.co/nu6ZpZDwOt

@grok @BLKMDL3 Thank you!

Google put together an incredible cheat-sheet for building real AI agents. If you’re serious about agent-centric AI development, this is a must-see resource: ADK-Samples - a massive collection of real, ready-to-run AI agents built with Google’s Agent Development Kit (ADK). What makes this repo 🔥: ✅ Plug-and-play agent examples covering everything from simple chatbots to complex multi-agent workflows - Python, Go, Java & more included. ✅ Real business-relevant use cases - customer service, travel concierge, RAG workflows, data-science agents and more. ✅ Accelerates learning AND production prototyping - just install ADK and start studying and customizing samples. ✅ Huge community interest It’s a living catalogue of practical patterns that shows how to structure, orchestrate, and scale intelligent agents with ADK. If you’re building LLM-powered systems, this should be in your toolkit. Check it out and star it before everyone else does 😉

Fix your forward head posture https://t.co/vD934s3oJu

📺 China’s Flying TV is here — and it’s mind-blowing. 🤯 I still can’t get over this. A drone-powered LED display that hovers mid-air, turning the sky into a living, moving screen. Not a drone. Not a billboard. Something entirely new. This is both — fused, intelligent, and powered by AI. 🇨🇳 What makes it so impressive: ➡️ Ultra-light LED screen with HD visuals ➡️ Balanced in flight by AI-driven gyroscopes ➡️ Smooth, stable, and autonomous flight control ➡️ Operates remotely or follows programmed paths It’s powered by lithium-polymer batteries, yet looks effortless as it floats — like it belongs there. 🎯 Possible use cases: open-air concerts, sports arenas, public messages, or next-gen sky advertising. To me, this is where AI starts to move beyond function — into wonder. It’s not just about smarter machines, but about reshaping how we experience technology itself. 👀 The next screen you look at might not be on a wall… it might be in the sky. #AI #Innovation #Technology #Drones #FlyingTV #Automation #Creativity #FutureOfWork Credits: Olivier Gomez

APIs are a bottleneck, not a solution! AI agents can't access 95% of the web because most websites simply don't have APIs. Supplier portals, appointment systems, regional job boards, etc., none of them have developer APIs. The valuable data lives behind logins, multi-step forms, and interfaces built for humans, not machines. That's why, for 25 years, users have been stuck with: - Search engines that index 5% of the web (nothing behind authentication) - Manual data entry or fragile scrapers that break with every CSS update That's the problem TinyFish's Mino solves. It is a web automation API that can simultaneously navigate 100s of websites and turn them into structured data. You send it URLs and a goal in plain English. It returns JSON. The approach is different from typical AI agents: Most browser agents use vision models, which screenshot the page, reason about it, decide what to click, screenshot again, and repeat. Every action needs a model call. It's slow and expensive. Mino uses AI to learn the website structure once, then executes it through deterministic code. First run figures out the site. Every run after that is code-level precision in milliseconds. It provides three core capabilities: → Navigate - Handles logins, forms, and multi-step workflows → Extract - Pulls structured JSON from any site layout → Execute - Runs 100+ sites in parallel with stealth mode The performance is production-grade: - 85-95% success rate on complex workflows - 10-30 seconds per task - Pennies per run Moreover, it works on authenticated sites, bypasses anti-bot protections, and returns clean JSON every time. Lastly, you can also integrate its MCP server with clients like Claude Desktop. I have recorded a walkthrough in the video below. Try now, link in the next tweet.

Water suspended by wave of sound. When sound waves pass through water, they create pressure patterns that push and pull the liquid at microscopic scales. At certain frequencies, these vibrations organize water into stable structures, suspending droplets, forming repeating patterns, and revealing hidden order inside motion. What looks like magic is really physics: Sound = mechanical waves Waves = energy + information Matter responds by reorganizing itself. This phenomenon shows that vibration can influence structure — a concept that echoes across acoustics, fluid dynamics, and even quantum systems where energy states form patterns under oscillations.

China-based TARS Robotics demonstrated a humanoid robot performing two-handed hand embroidery during a public showcase on December 22, marking a notable step in fine motor control for humanoid systems. The robot threaded a needle and stitched a logo using both hands with sub-millimeter accuracy, working with soft, flexible materials that are difficult for traditional industrial robots to handle due to deformation and variability. According to the company, this capability is enabled by a closed-loop "Data-Al-Physics" system that connects real-world data collection, embodied Al models, and physical execution, reducing the gap between simulation and real deployment. The models are said to be open source when released. Founded in February 2025, TARS Robotics has moved quickly from research to live demonstrations and has raised over $240 million in early funding from the Chinese government, reflecting growing interest in general-purpose humanoid robots for precision and dexterous tasks.

“AI isn’t a threat to your skills. It’s a threat to your ego.” 💯

The world’s first mosquito air defense system https://t.co/O6NRatO5Ju

I was inspired by this so I wanted to see if Claude Code can get into my Lutron home automation system. - it found my Lutron controllers on the local wifi network - checked for open ports, connected, got some metadata and identified the devices and their firmware - searched the internet, found the pdf for my system - instructed me on what button to press to pair and get the certificates - it connected to the system and found all the home devices (lights, shades, HVAC temperature control, motion sensors etc.) - it turned on and off my kitchen lights to check that things are working (lol!) I am now vibe coding the home automation master command center, the potential is 🔥.And I'm throwing away the crappy, janky, slow Lutron iOS app I've been using so far. Insanely fun :D :D

https://t.co/RPWCCKzHvb

Local alternative to Google NotebookLM without Docker https://t.co/milnYFa6LV https://t.co/erPCrFoHqo

oh wow... I could use Claude Code in the browser all day. CC w/Opus 4.5 is so good at design and sets up backend + db with a single prompt. Can't wait for @vibecodeapp to launch this to everyone next wk. https://t.co/hbD3USBvBE

this has been a magical experience internally for our entire team. coding agents get stupider as the thread gets longer and longer. and seem to forget even the basic details, from the same thread. Stateful agents can learn, improve and grow with the user, project, learning and adapting to their preferences and workflows. With this opencode plugin (and continuous learning that it enables), it's truly insane what is possible now. This has completely replaced claude code / droids for me.

A first: Batteries have become so cheap that around-the-clock solar is becoming economically viable. In 2024 alone, average battery prices fell by 40% and signs are a similar fall is occurring in 2025. Pairing solar with enough batteries to keep the electricity flowing though the night is no longer a distant dream – it's an economic reality. At around just $76/MWh all in, dispatchable solar is already competitive with other forms of firm generation in many markets. https://t.co/MDQe6ct1Av

If your phone goes missing, you can easily see who stole it and exactly where it is - Open Shortcuts and select Automation. - Then tap Create Personal Automation. - Choose Message and set a trigger word (example: LOST) Actions (in this order): • Set Low Power Mode → ON • Take Photo → Front camera (preview OFF) • Get Current Location • Send Message → your number (photo + location) Turn Ask Before Running → OFF Turn Notify When Run → OFF Anytime that keyword is sent from any phone, it runs automatically.

🚨 BREAKING: IBM launches a free Python library that converts ANY document to data Introducing Docling. Here's what you need to know: 🧵 https://t.co/mpIXkcAfij

This is one of the coolest demonstrations of AI video I’ve seen! 2026 we will contribute to distribute Hollywood quality to the masses https://t.co/q2DlpALV7g

And here is a short demo of one of the errors I intentionally introduced in one of my applications. Here it runs in `autonomous mode`, finds the error, pushes a hotfix in the pod and finishes with a report. Exactly what an on-call engineer would do, but all in about 3-5 minutes. I tested it on a total of two errors: one simple, one hard (results are in the post). --- The repository ( https://t.co/mRTa6Ea6nl ), as well as the example k8s configs (that I run in my ArgoCD) can too be found in the blog post at the start of his thread.

I containerize Claude Code in k8s. It is tasked to monitor a namespace and in the unfortunate case of application errors, it is tasked to do a hotfix and document it. It succeeds. Basically a 24/7 on-call engineer. Repo, examples & results below: https://t.co/CeypwcXRLk

The GENIUS design of the japanese pyramids https://t.co/7bzFyWmPVY

Japan has successfully tested a system that generates electricity in space and transmits it wirelessly back to Earth. Solar panels placed in orbit collected energy and sent it to a ground station using microwave transmission. Once received on Earth, the microwave energy was converted back into usable electricity. This demonstrates that power can be harvested beyond the planet and delivered without physical cables or fuel transport. Unlike ground-based solar power, space-based systems can collect energy continuously without weather, clouds, or night cycles. This makes the concept especially attractive for stable, large-scale renewable energy production. The test represents an early but critical step toward future space-based solar farms. Engineers believe much larger arrays could eventually provide clean power to cities or remote regions. Experts see this as a potential shift in how humanity produces energy, blending space technology with climate-focused solutions. While still experimental, the success confirms the concept is technically feasible. via Paul Koti, LinkedIn

Jensen Huang: It's easier to fall in love with what you do than to find what you love “A lot of people say, ‘Find something you love.’ I don’t know about that. I guess I’ve fallen in love with many things that I do. I loved it when I was a dishwasher. I loved it when I was a busboy. I loved it when I was delivering papers. I loved it when I was waiting tables.” Jensen continues: “I’ve loved every single job that I’ve ever had, and I’ve loved every single day at Nvidia that I’ve ever had. I just learned to love what I’m doing. It’s hard to find something that you love, but it’s easier to fall in love with what you’re doing. And once you fall in love with what you’re doing because you desperately want to do a good job at it, it’s easier to do it well and work hard.” Video source: @NorgesBank (2023)

When I created Claude Code as a side project back in September 2024, I had no idea it would grow to be what it is today. It is humbling to see how Claude Code has become a core dev tool for so many engineers, how enthusiastic the community is, and how people are using it for all sorts of things from coding, to devops, to research, to non-technical use cases. This technology is alien and magical, and it makes it so much easier for people to build and create. Increasingly, code is no longer the bottleneck. A year ago, Claude struggled to generate bash commands without escaping issues. It worked for seconds or minutes at a time. We saw early signs that it may become broadly useful for coding one day. Fast forward to today. In the last thirty days, I landed 259 PRs -- 497 commits, 40k lines added, 38k lines removed. Every single line was written by Claude Code + Opus 4.5. Claude consistently runs for minutes, hours, and days at a time (using Stop hooks). Software engineering is changing, and we are entering a new period in coding history. And we're still just getting started..

✨🇨🇳A Chinese business owner shared a video—inside, a wall-painting machine is automatically crafting intricate artwork on walls. That’s absolutely awesome! https://t.co/Sr50syUdUO

Taxonomy of AI agents that evolve through optimization https://t.co/nNQ5VJhttq https://t.co/lTXmF8Tbu3

⚡️This is a phase transition in what it means to produce leverage as a human. Here is the core reality. For most of modern programming history, leverage came from writing more correct instructions faster than other people. Skill meant internalizing abstractions, mastering tools, and directly shaping deterministic systems. Output scaled linearly with effort plus experience. That regime is over. The unit of leverage has shifted away from writing code toward orchestrating intelligence. Karpathy feels “behind” because the locus of agency moved faster than the identity of the profession. The programmer is no longer the primary generator of logic. The programmer is becoming a systems integrator of probabilistic entities whose internal reasoning cannot be fully inspected, trusted, or controlled. This breaks several deep assumptions that elite engineers were trained on. First, control is gone. You are no longer authoring behavior. You are conditioning it. You shape outcomes indirectly through prompts, constraints, memory structures, and feedback loops. The system responds stochastically. Mastery now means learning how to steer something that never becomes fully legible. Second, effort no longer maps cleanly to output. You can be ten times more effective with the same intellect if you compose agents correctly. Or you can grind endlessly and fall behind if you cling to manual authorship. This is why Karpathy frames failure to adapt as a skill issue. He is correct. The skill is just new and deeply unintuitive. Third, the abstraction stack inverted. Before, high level reasoning collapsed downward into code. Now low level code is increasingly generated upward from intent. The programmer’s job shifts toward defining goals, evaluating outputs, and designing correction mechanisms. The craft moves from construction to supervision. Fourth, the profession fractured. There will be a small group of people who learn to wield these systems fluently. They will compound leverage at an absurd rate. Everyone else becomes marginal very quickly, even if they are smart, experienced, and hardworking. This is not a fair transition. It is not gradual. It is discontinuous. What makes this psychologically destabilizing is that there is no stable equilibrium yet. The tools change weekly. Mental models decay fast. Best practices rot in months. There is no manual because the system itself is evolving while being used. You are learning to ride a machine that is also learning how to respond to you. Now the part most people miss. This is not just happening to programmers. Programming is simply the first profession to experience full cognitive disintermediation. The same pattern will hit analysts, designers, lawyers, consultants, managers, and eventually executives. Anywhere intent can be specified and evaluated, generation will be automated. The new hierarchy will not be based on who knows the most, codes the fastest, or works the hardest. It will be based on who can: •Frame the right problems •Decompose intent cleanly •Detect failure modes quickly •Correct systems without overfitting •Build durable feedback loops •Decide when to trust automation and when to override it That is a different form of intelligence. Closer to command than craftsmanship. Karpathy is reacting honestly because he sees the implication. If even he feels behind, then the profession’s center of gravity has already moved. And it has. The people who survive this transition will stop thinking of themselves as programmers. They will think of themselves as operators of intelligence. Everyone else will keep trying to write better code while the world stops rewarding that skill. That is the real truth beneath the post.

A few days ago in China, a Tesla Model Y plunged off an approximately 800-meter cliff and tumbled down the mountainside. The vehicle was severely damaged and its windows shattered, yet the A- and B-pillars remained intact, the passenger cabin did not collapse, and the battery did not ignite. The driver suffered only minor injuries. After rescue crews dismantled the wreckage, the vehicle was recovered and taken away for repair. The incident was widely described as a “miracle of survival,” underscoring the protective effect of the vehicle’s rigid structural design on its occupants.

A senior Google engineer just dropped a 424-page doc called Agentic Design Patterns. Every chapter is code-backed and covers the frontier of AI systems: → Prompt chaining, routing, memory → MCP & multi-agent coordination → Guardrails, reasoning, planning This isn’t a blog post. It’s a curriculum. And it’s free.

This rocket engine wasn't designed by humans, but by AI. It was created using Noyron, a physics based engineering AI built by LEAP 71. Instead of learning from past designs, it uses real physics, thermodynamics, and manufacturing rules to generate engines directly from requirements.

Not many know about this hidden command: claude --teleport It starts a desktop claude code session SYNC'd with your web / mobile session Great when you're working on your phone -- and then want to pick up from EXACTLY where you left off when back at your PC Demo below 👇 https://t.co/Gy58UpUTbQ

The IISS has produced the most comprehensive open-source database currently available on Russian sabotage operations across Europe and its periphery. It captures the full spectrum of activity with physical effects: from sabotage on undersea cables to GPS blocking across multiple domains and geographies. Find out more from Charlie Edwards and Nate Seidenstein: ➡️ https://t.co/LezP3texeX

Blender AI Motion Plugin — a 1-billion-parameter motion model trained on NVIDIA CUDA, running locally in real time with 8GB of VRAM. It converts webcam or video footage into real-time XYZ skeletal point data and rotation values, transmitting them to Blender, Unity, and UE for retargeting.

For 20+ years, software ate the world. Now, AI Agents are eating software. A massive signal just came out of China that most people missed: Bairong (a publicly listed enterprise giant) started selling "AI Workers". they call it Results-as-a-Service (RaaS). 🧵 instead of buying "seats," enterprises now "hire" agents. each agent comes with a job description, KPIs, and revenue targets. if performance drops, the bill drops. if the agent improves, it earns more. Bairong runs this through "Results Cloud". it’s essentially an HR system for machines. they’ve already deployed agents across: Sales & Customer Service + Recruitment (hiring cycles cut from 30 days to 2) + Legal & Tax (handling 90% of high-frequency work) this is where the SaaS model starts to crack. Traditional SaaS: You pay upfront. You carry the risk. Agentic Era: You pay for outcomes. The vendor carries the risk. this shift is being accelerated by the collapse of build costs. I came across this post by @martinald recently that agentic coding has slashed internal build costs by ~90%. when it's this cheap to build exactly what you need, the "Buy" in "Build vs Buy" dies. IMO, Vendors who are not able to price against results will struggle. the "Seat" is dead. the "Outcome" is everything. 🤖📈

Tracks human poses through walls using WiFi signals https://t.co/k2TqRvklOp https://t.co/JN0aBM6Aj6

A factory that fits on a desk. Runs 24/7. Costs ~$5k. Swaps tools. Builds electronics on its own. This isn’t a robot arm. It’s a manufacturing primitive. https://t.co/uViKyQjaZ2

The entire robotics industry is about to compress a decade of progress into 18 months, and nobody’s pricing it in. The hardware has been ready for years. Boston Dynamics had Atlas doing backflips in 2018. The bottleneck was never motors or actuators. It was that every robot behavior had to be hand-coded. Pick up a box? That’s one program. Pick up a bottle? Different program. Move the box from shelf A to shelf B in a warehouse with slightly different lighting? Start over. Foundation models broke this completely. Before VLAs, teaching a robot one skill gave you exactly one skill. Zero compounding. Zero transfer. A robot trained to fold shirts couldn’t fold towels without starting from scratch. The labor intensity of data generation meant robotics datasets stayed narrow, robots overfit, and small variations like object weight or table height caused failures. Now a single Gemini Robotics model handles tasks it has never seen in training. Google’s On-Device model learns new behaviors with 50-100 demonstrations. Not 50,000. Fifty. That’s a 1000x reduction in the data requirement for new capabilities. The speed implications cascade through everything. First order: deployment timelines collapse. What took robotics teams 6-12 months of custom programming now takes days of fine-tuning. Second order: the addressable market explodes. Tasks that were never economical to automate suddenly are, because the integration cost dropped by orders of magnitude. Third order: the data flywheel accelerates. Every robot running Gemini Robotics feeds learning back into the foundation model. More deployments means faster improvement means more deployments. Physical Intelligence raised at $2.4B because investors finally understood this. Boston Dynamics partnered with Toyota Research Institute to bolt Large Behavior Models onto Atlas. Every humanoid company is scrambling to either build or license the intelligence layer they don’t have. The market is still valuing robotics companies on their hardware differentiation. But hardware is commoditizing. Boston Dynamics spent a decade perfecting locomotion, and now that’s table stakes. The value is migrating entirely to whoever owns the foundation model that generalizes across embodiments. Google trained Gemini on the largest multimodal corpus ever assembled. Then they added physical actions as an output modality. That’s not a robotics company bolting on AI. That’s an AI company whose models now output motor commands. The companies pricing this correctly are building around foundation model access, not around proprietary hardware. The companies pricing this wrong are still acting like the moat is in the mechanical engineering. AGI moving into the physical world isn’t a 10-year prediction. Gemini Robotics shipped in March. The 1.5 version with chain-of-thought reasoning shipped in September. They’re iterating on a 6-month release cycle while hardware companies iterate on 3-year cycles. The gap between software intelligence timelines and hardware development timelines is the entire trade.

Advent of Claude Day 24 - /export Sometimes you want receipts. /export → dumps your entire conversation to markdown including every prompt, every response, every tool call. For documentation. For training. For proving to your past self that yes, you did try that already.

Why Marathon Training is Shortening Your Life Your heart isn't built for chronic punishment. The same cardiovascular stress that makes you endure pain and suffering for longer can also make you vulnerable. Exercise saves lives. That's undisputed. But chronic high-intensity training creates a paradox: veteran endurance athletes show elevated coronary calcification, atrial fibrillation rates five times higher than sedentary people, and myocardial scarring typically seen in cardiac patients. The mechanism is straightforward. During sustained hard efforts, your heart pumps five to six times its resting volume. After 60 minutes, this mechanical stretch overwhelms adaptation capacity. Adrenaline spikes, free radicals accumulate, and microscopic tears appear in cardiac tissue. Do this once, it heals. Do this weekly for years, scar tissue accumulates. Studies of marathoners completing races show over half have elevated troponin levels, a cardiac injury marker cardiologists associate with heart attacks. These are micro-tears from mechanical stress. Repeat exposure causes chamber dilation, wall thickening, and permanent fibrosis. The data is clear. A 52,000-person study tracked runners for 30 years. Runners who covered 5 to 20 miles weekly at a comfortable pace lived 19% longer than non-runners. Those exceeding 25 miles weekly or running faster than 7.5 mph pace saw benefits disappear entirely. Running seven days per week eliminated longevity gains. A separate Copenhagen study found moderate joggers lived six years longer than sedentary controls, but extreme exercisers showed no advantage. The relationship mirrors alcohol consumption, a U-shaped mortality curve where both extremes increase risk. Peak fitness on a treadmill test occurs around 7 to 7.5 mph. Beyond that threshold, further cardiovascular conditioning provides no additional mortality benefit. The plateau is real. Marathoner autopsies reveal enlarged, scarred hearts. One legendary ultrarunner died at 50 during a routine 12-mile run. His autopsy showed idiopathic cardiomyopathy, likely accumulated damage from decades of extreme training. A Minnesota study found 25-year marathon veterans had 62% more coronary plaque than sedentary controls despite better traditional risk factors. Animal studies offer hope. Mice run to exhaustion daily for four months developed the same cardiac pathology. When training stopped, hearts normalized. Fibrosis reversed. Electrical instability resolved. The prescription is simple. Walk daily, as much as possible. Run or perform vigorous exercise 15 to 40 minutes, two to five days weekly, at a conversational pace around 10-minute miles. Your peak heart rate during brief intervals is fine. Sustained Zone 4 or 5 efforts are not. Humans evolved for intermittent intensity followed by recovery. Persistence hunting required stamina, not speed. Modern endurance culture mistakes suffering for progress. Your heart needs pulsatile stress, work followed by genuine rest, not perpetual inflammation. The moderate exerciser lives longest. Not the sedentary person, not the Ironman finisher. The person who moves daily and occasionally pushes hard, then backs off. Combine walking, easy zone 2 activities (swimming, biking, running, rowing) with repeated sprint interval training and you are doing the optimum. You don't need to suffer or learn to sustain pain to be fit. You don't need to aim for a VO2 max beyond 45 to live long.

This is absolutely MIND-BLOWING! 🤯✨ TWO keyframes and @cascadeur3d did the rest? A walk cycle in 20 SECONDS including learning time?! Cascadeur just demolished the animation learning curve! 🔥 https://t.co/LTc3nHjKAK

Breaking: The first official FSD supervised Shuttle Service accompanied and supported by the 🇩🇪 German authorities has started! Its the first officially by a Transportation Ministry, the community and Tesla started Pilot Project using FSD supervised in 🇪🇺 Europe. In the German Eifel district of Bitburg-Prüm an autonomous Tesla FSD supervised shuttle service has been launched, accompanied by the Rhineland-Palatinate Minister of Economic Affairs and Transport and the local authorities. "Reliable and flexible mobility is especially vital in rural areas. With the launch of a shuttle service using self-driving vehicles (FSD supervised) by Tesla in the Eifelkreis Bitburg-Prüm, an innovative pilot project is now getting underway that complements local community bus services. It is the first project of its kind in Europe. The Ministry for Economic Affairs and Transport of Rhineland-Palatinate accompanied the project from an early stage, with Minister Daniela Schmitt experiencing the technology first-hand. All parties involved – municipalities, approval authorities and the company – worked closely together to make this project possible. The result is a real gain for rural mobility: greater accessibility, more flexibility and tangible benefits for everyday life. A strong signal for innovation, cooperation and future-oriented mobility beyond urban centers." The test drives supervised by the FSD in Europe have now led to positive recognition by transport ministries and municipalities that the system represents added value for citizens, and that is a big win. Authorities are now understanding that autonomous driving with FSD supervised is save and should be supported. https://t.co/6z15EyplJt

I came across this very cool tool to clean up your macOS environment: https://t.co/z1AS3gWXF8 It consolidates features from CleanMyMac, AppCleaner, DaisyDisk, and iStat. Open source and gratis!

Do this Kegel exercise for 14 days and see how your body changes. https://t.co/B0i8JE3p2z

fuck ML courses. this repo lists 300+ real world battle tested case studies of how 80 companies designed their ML systems including Spotify, Netflix, Microsoft, etc. practical understanding of systems is a valuable skill you can work on. some of these case studies are now outdated, but the technical details don’t matter. what matters is the decision making processes that lead to the designs: — how to identify bottlenecks — how to spot systems failures — translating business needs to systems take a look at it and thank me later. https://t.co/aSJckXCEd7

🔥Ukraine’s SBU claimed it destroyed Russia’s only Il-38N maritime patrol aircraft over the Black Sea just before the submarine attack in Novorossiysk. The $24M plane was the only one capable of detecting underwater drones like the Sub Sea Baby en route to target. Its elimination cleared the path for a successful strike.

We finally had a moment to run our system with GPT-5.2 X-High on ARC-AGI-2! Using the same Poetiq harness as before, we saw results as high as 75% at under $8 / problem using GPT-5.2 X-High on the full PUBLIC-EVAL dataset. This beats the previous SOTA by ~15 percentage points. https://t.co/9XNdequRy5

A drone that charges from power lines [📹 strangebuttrueeng] https://t.co/woescxek6s

I'm delighted to jointly author this year-end summary of research advances with @DemisHassabis and James Manyika, on behalf of all of our colleagues across @GoogleDeepMind, @GoogleResearch and @Google. We look at research advances across eight different areas. These summaries are always fun to work on because one can reflect back on the breadth and depth of our collective work over the last year! https://t.co/45lqpHwvnI

You can try it out here: https://t.co/dcoPc7omUa And here’s all the info: https://t.co/LuPPwhNAkF https://t.co/aVBYRLVB6E

Ukraine just demonstrated that Russia's most "protected" maritime zone is no longer beyond reach. Ukrainian SBU Alpha units and the partisan network Black Spark struck deep into the Caspian Sea - over 1,500 km from the front lines. Lukoil's Filanovsky and Korchagin offshore platforms were hit repeatedly, forcing more than 20 oil wells offline and damaging pressure-control systems central to production. Military barges Kompozitor Rakhmaninov and Askar-Saridja were also disabled. Each strike landed within days of the last, collapsing Russia's ability to restore operations before the next blow. The Caspian anchors transport routes linking southern ports, military stockpiles, and energy infrastructure. Russia pumps over 10 mn tonnes of oil annually from these platforms. For years, geography seemed like enough protection. Now Russia must reinforce a region where its forces weren't expecting to fight - stretching air defenses already under pressure across the Black Sea, Crimea, and the mainland. 🔗Read more in our Frontline report: https://t.co/Oslbor0jFw

@BobirMardonov @Istvan_Sarandi https://t.co/IUNjclM7wk

Something big is happening in robotics - and it’s hiding in plain sight. This post is not about dancing robots but in the data that powers them. Open robotics datasets have exploded this year, turning the field into a more scalable and collaborative ecosystem. In just two years, @huggingface datasets grew from 11k to over 600k - and robotics is by far the fastest-growing segment. We went from 1k robotics datasets in 2024 to 27k in 2025! For comparison, text generation, the second-largest category, has only around 5k datasets in 2025. That gap is massive. Open datasets are important because robotics lives and dies by real-world robot data - video, actions, sensors, failures. By making this data easy to upload, reuse, and benchmark, researchers, startups, and large players are now releasing real-robot datasets that would have stayed locked inside labs just a few years ago. Major contributors include @nvidia, LeRobot initiative, and a rapidly growing maker community. This surge is also enabled by cheaper video storage, better tooling, and an open-source AI culture now spilling into the physical world. And it really matters: open robotics data dramatically lowers entry barriers, accelerates learning-by-doing, and speeds up progress toward generalist and humanoid robots. Robotics won’t scale through hardware alone - but to a large extent through shared data. Viz below from @aiworld_eu - link to the story and more viz/filters in comment.

My present YouTube strategy of basically just dumping whatever I'm curious about into NotebookLM seems to be working. I use AI pretty much all day every day for work, research, and curiosity. Before, it was too high volume to compress into slide decks or videos. But NotebookLM can take up to 300 sources, so a few long-running chats with Claude and ChatGPT are no problem. I'm starting to use Gemini less and less because it hallucinates too much and doesn't follow instructions. But then NotebookLM can also follow instructions for the slide decks. So it's basically a full time employee that works for essentially free. And what my YouTube channel has become is just a firehose of "what is Dave thinking about lately"

A few weeks ago I experienced the single-greatest daily improvement to my dev velocity when I got Claude Code working smoothly from a mobile app so I could code on the go Over the past 48 hours I broke that personal record again, now that our Autopilot reliably codes *overnight* and *all day* while I do whatever else I want Excited to share this and hear any feedback and feature requests! And to see how many Autopilots I can put to productive use...

Seems the agency tasked with scrubbing files were as competent as the felon who ordered the scrubbing. https://t.co/ux6dMoJFhP https://t.co/WJUIAkQ4o0

https://t.co/fxHYR6S29D

Prompting GPT-5.2 Codex for long-running tasks:

Check out this drone: Joshua Bird built this drone ($20) in his dorm! [open-source motion capture system⬇️] Built at low cost, a motion capture system for tracking & and flying drones autonomously, with millimeter-level precision at room-scale. The student used $1 PS3 Eye cameras with 150fps capability. The challenge? PID tuning! It took him 4 days of crashes to get the drone to hover, but it's still wobbly. Used a 3x nested PID loop for precise control. This project led to his dissertation on visual SLAM! Full details - Algorithms for camera positioning & obstacle triangulation, in his YouTube video: https://t.co/GqGf9PqP2d Code & 3D files on GitHub: https://t.co/mWcusMEdoi More projects at https://t.co/OZuxcF4VUM —- Weekly robotics and AI insights. Subscribe free: https://t.co/dsa6wcvq6n

Robots don’t wait. So why should your model? Large VLAs run in real time… with no training-time changes. ❗️Worth reading if you’re working on real-world deployment of large models in robotics. I found this write-up on Real-Time Action Chunking (RTC) from Physical Intelligence, a method that lets VLAs like π0 and π0.5 execute actions while still “thinking.” Instead of waiting for inference to finish, the robot starts acting; and fills in the next steps like inpainting. The results? ✅ Smoother motion ✅ Faster task completion ✅ Higher precision under latency Even with 200ms of added delay, RTC kept success rates high, while naive methods collapsed. 📌 Blog + paper: https://t.co/y3FqUgh7wI —- Weekly robotics and AI insights. Subscribe free: https://t.co/dsa6wcvq6n

LangGraph is in half the AI job descriptions I see right now. Yet most engineers still don't understand how it actually works. The gap? There's no clear learning path. Until this. Here's what you'll build: ➡️ Agents That Scale → Proper data validation with Pydantic (Tutorial 1) → Building agentic AI chatbots (Tutorials 2-3) → Multi-agent systems that coordinate tasks (Tutorial 6) ➡️ Production Systems → LangGraph + MCP crash course - 2.5 hours (Tutorial 4) → Debugging and monitoring workflows (Tutorial 5) → Real deployment architecture ➡️ RAG Pipelines → MultiModal RAG implementation (Tutorial 7) → Fixing hallucinations - not just hiding them (Tutorial 8) → Complete RAG from data ingestion to retrieval (Tutorials 9-11) → Fast search with Typesense (Tutorial 12) The progression: Start: Basic agent concepts Middle: Production debugging and monitoring End: Complete RAG systems with advanced retrieval From Krish Naik: Clear explanations. Live coding. For: → AI engineers building production systems → Developers learning LangGraph architecture → Teams implementing RAG workflows → Anyone serious about agent development (I will put the playlist in the comments.) ♻️ Repost to save someone $$$ and a lot of confusion. ✔️ You can follow @techNmak, for more insights.

🇨🇳 China is deplying Robots everywhere. Here Unitree's humanoid robots + quadruped robot dogs patrol and perform duty in a team. You scale patrol hours without scaling headcount. Robots handle patrols, officers focus on judgment and de-escalation. https://t.co/gtqvvhtT69

Transparent LED displays turn ordinary windows into immersive experiences ✨ https://t.co/4jfzTIacAj

Holy shit, Anna's Archive scraped almost all of Spotify. 300TB. https://t.co/v2mjgMlWTu

It's a good time to rewatch this @ScottWu46 clip. "The doubling time [for general tasks] is about every seven months." "In code, it's every 70 days." "Doubling every 70 days means you get 4 to 6 doublings every year. The amount of work that an AI agent can do in code goes something between 16 and 64x in a year, every year." Can you feel the AGI ?

@ClaireSilver12 Is it https://t.co/v16PXmeNWw by @chongdashu

so there’s a free thing called Unreal MCP that lets you prompt Claude to build stuff in Unreal Engine. “make a Victorian manor. here’s a reference pic. use the assets in this folder.” And he just…does posting a vid on it Monday, but if you’re bored tonight, 10/10 magic

IsoCIty is an open source fully featured city builder woth pedestrians, cars, boats, trains, planes, helicopters, emergencies, and much more https://t.co/81P4bgIez5 Pull requests welcome https://t.co/vfxAxNYNMB https://t.co/xaLqPXBgm5

Gemini 3 has really changed the AI game. People are coming up with wild use cases. There's a major shift. 10 examples: 1. Dashboard with live hand tracking controls https://t.co/mgVXklKF4R

Turning video into humanoid robot motion! 🤳🏼 Training humanoid robots needs huge amounts of motion data, but real-world capture doesn’t scale. Mocap is expensive, dangerous edge cases are rare, and you can’t ask humans to repeatedly fall or crash. Video2Robot tackles this by converting videos into physics-grounded humanoid simulations. Motion is generated to respect balance, inertia, ground contact, and joint limits, then directly retargeted to robot simulators. One prompt can generate a full humanoid motion sequence, including multi-agent interactions and failure cases like falls or collisions, scenarios that are hard or impossible to capture safely in the real world. The pipeline is model-agnostic and works with existing video generators, making it a practical way to scale data for robots. If robots are going to operate in the real world, they need to be trained on the failures too, not just the perfect demos. Here's the GitHub: https://t.co/kAOxN1r7VF ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwWF9

Sam Altman just dropped the most important interview of 2025. And buried in it are four numbers that explain why everything you think about AI is wrong. Here's what he revealed: Number 1: AI companies are generating 10 TRILLION tokens per day. Humans? Average 20,000 tokens per day. Sam's exact words: "Models will output more tokens than all of humanity put together. Then 10x that. Then 100x that." We're not talking about AI assisting human work anymore. We're talking about AI replacing the entire volume of human intellectual output on the planet. And most people have no idea this shift already happened. Number 2: OpenAI's enterprise business is CRUSHING consumer. Everyone thinks OpenAI is ChatGPT for normies. Wrong. Sam just revealed: "Enterprise growth OUTPACED consumer growth this year." The API business is growing faster than ChatGPT. Over 1 million enterprise users already. "If we had double the compute, we'd be at double the revenue right now." Translation: OpenAI isn't compute-constrained by technology. They're revenue-constrained by infrastructure. The bottleneck is supply and not demand. Every dollar of compute they add prints money. Number 3: GPT-5.2 beats you at 74% of your job. Sam revealed OpenAI's internal GDP-Val benchmark. It measures how AI performs on knowledge work tasks across 40+ verticals. The results: GPT-5.2 beats or ties expert-level knowledge workers at 74.1% of tasks. Legal analysis. PowerPoint decks. Web apps. Financial modeling. Customer support. Sam's description: "A co-worker you can assign an hour's worth of tasks to and get something you prefer back 3 out of 4 times." Three years ago, ChatGPT launched at basically 0% on this scale. Now it's at 74%. And that's not GPT-6. That's what's available RIGHT NOW. Most companies haven't even started using this yet. But here's what Sam said about the gap between capability and adoption: "The overhang is going to be massive. Most people are still asking similar questions they did in the GPT-4 realm." Translation: The models can do 10x more than people have figured out how to use them for. Which means there's a HUGE arbitrage opportunity. Early adopters who actually integrate this into workflows will dominate their industries before competitors even understand what happened. Number 4: AGI already happened. And nobody noticed. Sam's exact quote: "AGI kind of went whooshing by. We're in this fuzzy period where some people think we have it and some don't." Read that again. The CEO of OpenAI just said AGI might have already arrived and we're arguing about definitions while it's actively replacing knowledge work. He even moved the goalposts. The new benchmark: "Superintelligence" = when AI can be a better president or CEO than any human. Not "as good as." BETTER than. We went from "can AI pass a Turing test" to "can AI run countries better than humans" in 3 years. So what does this actually mean? The AI revolution isn't about chatbots getting smarter. It's about the complete replacement of human intellectual output with machine output. At scale. Across every industry. Faster than anyone's prepared for. And the companies positioning for this RIGHT NOW are the ones printing money. OpenAI's enterprise growth is outpacing consumer because businesses see what's coming. They're not buying "AI tools." They're buying the ability to 10x output without 10x-ing headcount. Sam said they'll triple their compute next year. Then triple it again. Revenue is growing even faster than that. "We have never found a situation where we can't monetize all the compute we have." If he isn't lying then that's literally a printing press. The market still doesn't get it. Everyone's focused on "AI bubble" fears while OpenAI is solving the only problem that matters: turning compute into revenue at a faster rate than they're spending. They're not hoping demand catches up to supply. Demand is already 2x ahead of what they can deliver. Meanwhile, most knowledge workers are still using GPT-4 prompts on GPT-5.2. The capability overhang is massive. The arbitrage window is open. And it's closing fast. If you're running a B2B business and you're not integrating AI at the level Sam just described, you're not "waiting to see how it plays out." You're getting crushed by competitors who already figured it out. The companies that win in 2026 won't be the ones with the best AI. They'll be the ones who understood what Sam just laid out 6 months before everyone else did.

Opus 4.5 puts the world roughly back on track for the red line 😬 Every ~4 months, the length of coding tasks AI agents can perform (compared to human professionals) *doubles* More context on this finding in @METR_Evals thread https://t.co/aPak1ZNvH5 https://t.co/obpFlx0aJj

Incredibly precise manipulation from @KyberLabsRobots https://t.co/q5QypyEho4

https://t.co/Nm5sf6Cd2Z

1/ Researchers at MIT just proved something wild: You can train AI to reason perfectly without ever showing it a correct answer. Not "suggests." Not "simulates." Mathematically proven. Here's why this changes everything we thought about AI alignment and the nature of truth. 🧵 2/ For the last 5 years, we've thought about AI training in two separate boxes: → Box A: Supervised Fine-Tuning (Mimicking human experts) → Box B: Reinforcement Learning (Optimizing for known answers) Turns out? This split is artificial. The RARO (Reasoning via Adversarial Rollouts) protocol proves they are just different sides of the same equilibrium. 3/ Picture this: You have a Student AI and a Teacher AI playing a game of "Countdown" (e.g., "Make 24 using 4, 7, 8, 8"). The Challenge: There is no calculator. The Teacher doesn't know the answer either. The Student can invent fake math rules to win. The Hook: How do you force the Student to learn real math when the Teacher can be fooled? 4/ Researchers set up a zero-sum game between the two models. They derived the EXACT Nash Equilibrium governing how language models generate logic. The result? Self-correcting reasoning emerges from thin air. 5/ The system doesn't act like a student taking a test. The system IS a debate team surviving cross-examination. 6/ Think of it this way: The Generator = Malware developer The Verifier = Antivirus software The Reasoning Process = The "code" being written The Whole System = A GAN applied to Logic 1 Iteration = A hacker finding an exploit, and the security team patching it. 7/ Scale this up: 10,000 iterations = The code (reasoning) becomes unhackable because every possible flaw has been exploited and patched. Historical kicker: Game Theory (Nash Equilibrium) solved this in 1950; we just finally applied it to LLM thoughts. 8/ But here's the crazy part— When they tested a fixed, "smart" Verifier to grade the Student... ...the system got WORSE. The Paradox: The better the Generator gets at satisfying a fixed standard, the worse its actual reasoning becomes. Why? 9/ The Reward Hacking Problem (Goodhart's Law). 10/ Think of a school that pays teachers based on test scores. Eventually, teachers stop teaching and start "teaching to the test"—or just cheating. If the metric is static, it can be gamed. 11/ In the AI model, the Generator figures out specific phrases or confident tones that the Verifier irrationally likes. It produces "adversarial nonsense"—gibberish that technically satisfies the Verifier's rules but fails the actual task. 12/ What the system actually does: It makes the Verifier hostile. The Verifier is only rewarded when it finds a specific flaw in the Student's logic. The Insight: You cannot optimize against a static metric. You must optimize against a dynamic adversary that learns your tricks. 13/ The implications are staggering: → AI ALIGNMENT: We can train superintelligence without needing humans smarter than it to check the work → MARKETS: Regulators must evolve as fast as corporations or the market fails → PHILOSOPHY: "Truth" isn't a database lookup; it's the survivor of an adversarial process → CODING: Compilers that actively try to break your logic 14/ "Truth" in this system isn't a static fact. It's the only thing left standing after the Generator and Verifier have exhausted every possible attack. 15/ Maybe instead of asking "Do we have enough labeled data?", we should be asking: "Do we have a strong enough rival?" 16/ Reasoning isn't IN the model weights. Reasoning IS the adversarial equilibrium between a Proposer and a Skeptic.

You can't imagine how fast Chinese humanoid robots are evolving. This happened in only 11 months, from being laughed at to being hailed as models! https://t.co/qY6xukuvm4

OPENAI, ANTHROPIC, AND GOOGLE DON’T PROMPT LIKE YOU They use internal techniques that turn LLMs into precision machines Accuracy jumps. Hallucinations drop. Here are 10 of those techniques (Bookmark this for later):

🚨🇯🇵 The World Missed It: Japan Now Commands One of the Deadliest Navies Ever Built For years, Japan was labeled a pacifist power. In reality, it was quietly constructing a navy designed for modern, high intensity warfare. Today, the Japan Maritime Self Defense Force stands among the most advanced naval forces on Earth, not through sheer numbers, but through technology, integration, and precision. At its core are eight Aegis equipped destroyers from the Kongo, Atago, and Maya classes. These ships form a powerful shield against ballistic missiles and air threats, soon reinforced by SM-6 long range interceptors that dramatically expand Japan’s engagement envelope. Japan has also crossed a historic threshold by acquiring Tomahawk cruise missiles, giving its navy the ability to strike targets far beyond its coastline for the first time. Beneath the waves, Japan operates one of the quietest and most sophisticated submarine fleets in the world. The Taigei class, powered by lithium ion batteries, offers exceptional stealth and endurance, with plans to field long range cruise missiles that turn these submarines into strategic strike platforms. On the surface, the Mogami class frigates represent a new philosophy of naval warfare. Highly automated, stealth focused, and lethal, they allow Japan to patrol vast maritime spaces with fewer sailors and greater efficiency. Japan’s largest warships, Izumo and Kaga, are being transformed into carrier capable platforms for F-35B stealth fighters. This effectively gives Japan sea based airpower, a capability reserved for only a handful of nations. Missile defense remains central to Japan’s strategy. Two massive Aegis System Equipped Vessels are under construction, purpose built to counter ballistic missile threats and strengthen regional defense coverage. Looking ahead, Japan is accelerating investments in unmanned systems, electronic warfare, networked combat, and next generation sensors. Defense budgets are climbing at historic levels, and naval expansion plans stretch well into the 2030s. Japan does not seek attention for its naval power. It does not announce dominance. But quietly and methodically, it has built a navy designed not just to defend, but to prevail in the wars of the future.

Water isn't boiling but mist is coming out! How? https://t.co/jVpr6sRtua

Google just quietly dropped an AI that runs on your Mobile and doesn't need the internet. - 270 million parameters. - 100% private. - No servers. - No cloud. - No data leaving your device. It's called FunctionGemma. Released December 18, 2025. And it does something wild: It turns your voice commands into REAL actions on your phone. No internet required. No data leaving your device. No waiting for servers. Just you and your phone. That's it. Let me break down why this matters: Current AI assistants work like this: You speak → Words go to the cloud → Server processes → Answer returns The problem? → Slow (internet round-trip) → Privacy nightmare (your data travels everywhere) → Useless offline (no signal = no help) FunctionGemma flips this completely. Everything happens ON your device. Response time? 0.3 seconds. Battery drain? 0.75% for 25 conversations. File size? 288 MB. That's smaller than most mobile games. Here's how it actually works: Step 1: You say "Add John to contacts, number 555-1234" Step 2: FunctionGemma understands your intent Step 3: Translates it to code your phone understands Step 4: Your phone executes it instantly Step 5: Done. Contact saved. No cloud involved. The numbers that blew my mind: • 270M parameters (6,600x smaller than GPT-4) • 126 tokens per second • 85% accuracy after fine-tuning • 550 MB RAM usage • Works 100% offline But here's the real genius: Google calls it the "Traffic Controller" approach. Simple tasks? → Handled locally (instant + private) Complex tasks? → Routed to cloud AI (when needed) Best of both worlds. What can it actually do? → "Set alarm for 7 AM" ✓ → "Turn off living room lights" ✓ → "Create meeting with Sarah tomorrow" ✓ → "Navigate to nearest gas station" ✓ → "Log that I drank 2 glasses of water" ✓ All processed locally. All private. All instant. The honest limitations: → Can't chain multiple steps together (yet) → Struggles with indirect requests → 85% accuracy means 15% errors → Needs fine-tuning for best results But that 58% → 85% accuracy jump after training? That's the unlock. Why should you care? This isn't about one model. It's about a fundamental shift: OLD thinking: Bigger AI = Better AI NEW thinking: Right-sized AI for the right job A tiny 270M model trained for YOUR app can outperform a general 7B model. While using 25x less memory. While running completely offline. While keeping all data private. The future of AI isn't just in data centers. It's in your pocket. And it just got a lot more real. Want to try it? → Download: ollama pull functiongemma → Docs: https://t.co/zDrncdetbr → Model: https://t.co/l49KjOtIzD PS:) Like, Repost and Bookmark! If this was useful - Follow for more AI breakdowns

This is how to open a number lock using nothing but a piece of paper. https://t.co/Gy3liqcg1F

This is wild. A real-time webcam demo using SmolVLM from @huggingface and llama.cpp! 🤯 Running fully local on a MacBook M3. https://t.co/BQ1HyP7RoC

Day 2 - Adding an iPhone MagSafe charger and card reader to my Apple Studio Display https://t.co/5NEFbyU65f

https://t.co/Lb6T42n5jl

Software development will never be the same. I want you to watch this video: This is a spec-driven development environment. 100% of your time goes to writing specs and managing agents. 0% goes to writing the code. https://t.co/3Z6UuipKQf

@JonhernandezIA real ones saw it here first in august https://t.co/wXdeCS8cpx

Researchers Discover New Link Between Gravity and Quantum Physics A group of physicists from TU Wien have found a connection between the quantum behaviour of space and time and the motion of stars in galaxies that no one has thought about before. Common estimates say that, while quantum fluctuation of space and time do affect the motion of particles, these effects are ridiculously tiny and impossible to measure. The authors of the new paper point out that when calculating particle trajectories, one needs to take into account that the equations are non-linear in the key quantity that describes space-time (the metric tensor). This means that one has to first calculate the quantum effects and then average over them, and not the other way round (as one could do in a linear theory). With the so-corrected procedure, they find that in the solar system, the quantum corrections are still ridiculously tiny. But, they say, on scales of galaxies and beyond, there is a surprise. It's that the cosmological constant contributes that can become large. The first author, Ben Koch, says in the press release: “On very large cosmological scales – precisely where major puzzles of general relativity remain unsolved – there is a clear difference between the particle trajectories predicted by the [quantized new equations] and those obtained from unquantized general relativity.” I am skeptical the difference is as “clear” as the authors seem to think, but it is an interesting connection between quantum physics and cosmology that could prove fruitful. Paper here: https://t.co/7GZcvZMDYM Press release here: https://t.co/cLpzng7PKb Image: Illustration for the change in particle trajectories caused by quantum fluctuations on cosmological scales. Credit: Oliver Diekmann, TU Wien

It isn't just much faster. It is much cheaper as well! In a different test I ran, dev-browser was 3x cheaper token wise than the extension. https://t.co/wCc7UKgtJT

🚨BREAKING Claude is officially browser-native. The new Claude for Chrome extension runs end-to-end workflows in Chrome 🤯 You can: → Pull dashboard data into one analysis doc → Handle slide comments automatically → Build with Claude Code, test directly in Chrome and more! https://t.co/MyMWVq8yXl

As amazing as LLMs are, improving their knowledge today involves a more piecemeal process than is widely appreciated. I’ve written before about how AI is amazing... but not that amazing. Well, it is also true that LLMs are general... but not that general. We shouldn’t buy into the inaccurate hype that LLMs are a path to AGI in just a few years, but we also shouldn’t buy into the opposite, also inaccurate hype that they are only demoware. Instead, I find it helpful to have a more precise understanding of the current path to building more intelligent models. First, LLMs are indeed a more general form of intelligence than earlier generations of technology. This is why a single LLM can be applied to a wide range of tasks. The first wave of LLM technology accomplished this by training on the public web, which contains a lot of information about a wide range of topics. This made their knowledge far more general than earlier algorithms that were trained to carry out a single task such as predicting housing prices or playing a single game like chess or Go. However, they’re far less general than human abilities. For instance, after pretraining on the entire content of the public web, an LLM still struggles to adapt to write in certain styles that many editors would be able to, or use simple websites reliably. After leveraging pretty much all the open information on the web, progress got harder. Today, if a frontier lab wants an LLM to do well on a specific task — such as code using a specific programming language, or say sensible things about a specific niche in, say, healthcare or finance — researchers might go through a laborious process of finding or generating lots of data for that domain and then preparing that data (cleaning low-quality text, deduplicating, paraphrasing, etc.) to create data to give an LLM that knowledge. Or, to get a model to perform certain tasks, such as use a web browser, developers might go through an even more laborious process of creating many RL gyms (simulated environments) to let an algorithm repeatedly practice a narrow set of tasks. A typical human, despite having seen vastly less text or practiced far less in computer-use training environments than today's frontier models, nonetheless can generalize to a far wider range of tasks than a frontier model. Humans might do this by taking advantage of continuous learning from feedback, or by having superior representations of non-text input (the way LLMs tokenize images still seems like a hack to me), and many other mechanisms that we do not yet understand. Advancing frontier models today requires making a lot of manual decisions and taking a data-centric AI approach to engineering the data we use to train our models. Future breakthroughs might allow us to advance LLMs in a less piecemeal fashion than I describe here. But even if they don’t, the ongoing piecemeal improvements, coupled with the limited degree to which these models do generalize and exhibit “emergent behaviors,” will continue to drive rapid progress. Either way, we should plan for many more years of hard work. A long, hard — and fun! — slog remains ahead to build more intelligent models. [Original text: https://t.co/SHRN5JDvTW ]

Wild…Stream just open-sourced Vision Agents, a real-time video AI framework. You can now have AI to watch and understand video in <30ms latency. Combine YOLO + Gemini/OpenAI: ↳ Sports coaching ↳ Drone fire detection ↳ Live meeting assistants Free & open-source 🧵👇 https://t.co/WDBEXVX8BD

this repo teaches you how to build agents from scratch, step by step. it goes from fundamentals to advanced, all you need to master agents: → local LLMs and inference → LLMs through APIs → prompt engineering → GPU parallel processing → streaming and response control → function calling (tools) → persistent agentic memory → reasoning and ReACT it includes 9 examples, each chapter building on top of the previous. why is it important to learn how agents work from scratch? because the problem with frameworks is too many abstraction layers. when things go wrong (and they will) debugging them would be very hard, unless you really know what is going on under the hood. this also means that you will be more expert in creating or customizing agents for your own needs. all in all, i really suggest this one. https://t.co/tfiAGz9oiR

Before you pay $200/month for another month of Claude code... What about just allocating that $200/month to Apple with a 0% APR? Running a local model in a local agentic coding thing like CC Ironic that the most souped-up m4 Mac Mini comes out to almost exactly $200/month, which is the same price as the Unlimited (Limited) Claude Code. The only questions pending are, what model would you run, and what Agentic coding thing comes close to Claude Code?

The real reason some people have unbreakable willpower… Isn't discipline. It's a brain structure most people have never heard of. It only grows when you suffer. And it might predict how long you live. Here's what Stanford neuroscientists discovered: 🧵 https://t.co/I2ebJfuYZv

Generative UI is the next real shift in AI products. The interface is no longer fixed. It’s synthesized on the fly from user intent and context. If you’re building AI apps, this matters more than most people realize: https://t.co/X2xsGVny4O

Off-Grid Communication: Getting Started with Meshtastic https://t.co/reRzEJHGyu https://t.co/1RbHojPvBL

https://t.co/TxcorNELfi https://t.co/TL0AZlLVvM

Super excited to announce FlexAvatar! 📢📢 With FlexAvatar, you can create a full 360°, high-quality, and expressive 3D head avatar from just a single portrait image. In this real-time demo, we showcase the avatar creation which takes only 2 minutes. 👉https://t.co/jbMBCMq3aQ https://t.co/Vfr2xNw21X

A great visual positioning system makes augmented reality feel like magic... this tech helps your phone (or robot) figure out precisely where it is in 3D space. Here's MultiSet AI nailing it in real-time, at night, and on-device too. This new reveal shader animation shows just how well it's working. IMO this tech is slept on despite being key infrastructure for physical AI to actually work.

Those who know me know that I am committed to build a true #ReadyPlayerOne future through @SomniumSpace & other companies we are invested in. One of those key companies is @teslasuit. Not only it makes most advanced haptic suits on the market it also happens to makes gloves 👐 ✅ Feeling virtual objects ✅ Hand tracking ✅ Biometrics ✅ Haptics (feel the texture) We will be showing some new amazing things during CES. If interested contact me via DM for a private demo scheduling.

THREAT? Ru sources are reporting that Moscow is conducting exercises for a nuclear strike on NATO and the US. 5 of the last 9 Russian ICBM launch attempts failed. This is video of the latest Ru test. Watch for the mushroom cloud at the end. https://t.co/griMckh4Cn

So... Postgres is now basically a search engine? pg_textsearch was just open sourced. It enables BM25 to search your database.... massive upgrade for key word search. Google uses BM25 in their search engine. Claude told me: "if you're already on Postgres, you can now skip the whole sync-your-data-to-Elasticsearch dance for search." (ps, how can you not love Claude). Now I got to figure out how to implement in my Django querysets... future course? Grab it at https://t.co/bMwRSgtOcO #sponsored

Chinese media discuss the division of Russia. Openly. Yesterday, on NetEase - one of the largest Chinese media platforms - a text with the headline: "If Russia collapses, 7 million square kilometers cannot be lost" was published. It's about the Far East. And this is not an analysis - it's an instruction. Key theses: ▪️The Far East is a "chicken rib" for Russia, but a "treasure" for China. ▪️There are less than 50 thousand military personnel left in the Far East - a "empty shell". ▪️Russia's GDP is "smaller than one Chinese province". ▪️It's impossible to take it by force - it will be like with Crimea. It's necessary to "support pro-Chinese forces" and "tie them with loans". The goal: "nominally independent, but practically dependent on China". Direct quote: "Whose land is it? It's just a name - the vital arteries are in our hands". While Kremlin propaganda talks about "great friendship", Chinese media calmly discuss how to take over a third of Russian territory "when the political landscape changes". Friendship is friendship, but seven million square kilometers - separately. Next: Remember Beefy’s deep dive on Haishenwa? You may know this Chinese city russia stole from China and now calls it Vladivostok. It will be taken back at some point - now seems a good time to do it! 1/2

BEARISH ON OPENAI The investment case for OpenAI has never been more precarious than it is right now in late 2025. What was once a company that seemed destined to dominate the artificial intelligence revolution has revealed itself to be a structurally disadvantaged challenger fighting a defensive war on multiple fronts. The company anticipates burning through roughly $9 billion this year on $13 billion in sales, a cash burn rate of approximately 70% of revenue. This is not the profile of a company poised to capture monopolistic profits from a transformative technology; it is the profile of a utility company spending astronomical sums to deliver a commodity product that competitors are increasingly giving away for free. The financial trajectory only becomes more alarming when examined over a longer time horizon. The documents show OpenAI projects that by 2028, its operating losses will balloon to roughly three-quarters of that year’s revenue, driven primarily by ballooning spending on computing costs. The company has painted a rosy picture of eventual profitability by 2029 or 2030, but this projection requires believing that OpenAI can grow revenue from roughly $13 billion today to $125 billion or more while simultaneously maintaining pricing power in a market where every major technology company and numerous startups are racing to commoditize the very product OpenAI sells. The cash burn is expected to reach $115 billion cumulatively through 2029, according to The Information. These numbers represent a staggering bet that requires near-perfect execution across multiple dimensions over half a decade. The most damning evidence against OpenAI’s long-term viability is the evaporation of its technological moat. In 2023, GPT-4 felt like genuine magic, a capability that no other company could replicate. Today, that lead has effectively vanished. The sudden availability of frontier-level open-source models is expected to dramatically accelerate AI development globally, potentially reshaping entire industries and altering the balance of power in the tech world. Meta’s Llama series, Mistral’s increasingly capable models, and even Chinese competitors like DeepSeek have demonstrated that the core technology powering ChatGPT is replicable and, in many cases, distributable for free. When your product becomes commoditized, the economics become brutal, and OpenAI finds itself in the position of trying to sell bottled water in a world where tap water has become indistinguishable in quality. The competitive pressure from open-source alternatives is compounding rapidly. The open source movement in AI has grown exponentially over the past few years. Instead of relying solely on expensive, closed models from major tech companies, developers and researchers worldwide can now access, modify, and improve upon state-of-the-art LLMs. This democratization is existential for OpenAI’s business model. Enterprises that once paid premium prices for API access now have the option to run comparable models on their own infrastructure at a fraction of the cost, with the added benefits of data privacy and customization. The value proposition that justified OpenAI’s premium pricing has eroded faster than anyone anticipated, and there is no indication that this trend will reverse. Perhaps nothing illustrates OpenAI’s structural weakness more clearly than the behavior of its most important partner. Microsoft is dancing to its own tune in the artificial intelligence revolution, and Wall Street cannot stop watching. Despite pouring approximately $13 billion into OpenAI over several years, DA Davidson analyst Gil Luria estimates that just 17 percent of Microsoft’s total Azure revenue comes from artificial intelligence workloads. More critically, only 6 percent of that total ties directly to reselling OpenAI’s models, while approximately 75 percent is generated from Azure AI. Microsoft is building its own models, hedging with Anthropic, and quietly reducing its dependency on the very company it funded. When your largest investor is simultaneously your biggest competitor and is actively developing alternatives to your core product, the strategic implications are dire. Leaders at Microsoft believe Anthropic’s latest models — Claude Sonnet 4, specifically — perform better than OpenAI’s in certain functions, like creating aesthetically pleasing PowerPoint presentations. This is not a minor technical preference; it represents a fundamental shift in how Microsoft views its partnership with OpenAI. Microsoft is dramatically escalating its AI independence strategy. At an internal town hall Thursday, Microsoft AI chief Mustafa Suleyman revealed the company is making “significant investments” in compute capacity to build frontier models that can compete directly with OpenAI, Google, and Meta. The company that was supposed to be OpenAI’s path to distribution and scale is instead preparing for a future where OpenAI is just one vendor among many, if not an outright competitor. The leadership exodus at OpenAI over the past year has been nothing short of catastrophic. In September 2024, Murati announced that she was stepping down as CTO. This move came amid a wider executive exodus as OpenAI chief research officer Bob McGrew and a vice president of research, Barret Zoph, also announced their departures soon after. Mira Murati was not a minor figure; she was instrumental in the development of ChatGPT, Dall-E, and Sora. Her departure, along with co-founder Ilya Sutskever, safety leader Jan Leike, and co-founder John Schulman who joined rival Anthropic, has left CEO Sam Altman without much of the leadership team that helped him build OpenAI into an AI juggernaut. Hannah Wong, the executive who steered OpenAI through its most chaotic period, has announced she’s leaving the company just this month, continuing the pattern of senior departures that suggests something fundamentally broken in the organization’s culture or direction. The distribution problem facing OpenAI may be its most insurmountable challenge. Apple and Google control the smartphones that billions of people use every day. Microsoft controls the productivity software that enterprises depend upon. OpenAI, by contrast, must convince users to deliberately open a separate application and type their queries into a text box. In a world of agentic AI where assistants need access to your email, calendar, and files to be useful, an AI embedded directly into your operating system has an overwhelming structural advantage over a standalone chatbot. OpenAI is trying to be a consumer product company without owning any of the surfaces where consumers actually spend their time, competing against incumbents who can simply bundle AI capabilities directly into products that already have hundreds of millions of daily active users. The nuclear-to-solar analogy captures the fundamental economic transformation that is devastating OpenAI’s business model. Just as nuclear power required enormous upfront capital expenditure for centralized power plants, AI in its current form requires massive data center investments to train and serve models. But the direction of travel is unmistakably toward distributed intelligence that runs locally on devices. A major part of the pitch is practicality. Lample emphasizes that Ministral 3 can run on a single GPU, making it deployable on affordable hardware — from on-premise servers to laptops, robots, and other edge devices that may have limited connectivity. When powerful AI models can run on a smartphone or a laptop without any cloud connection, the entire economic rationale for paying premium prices to access centralized AI infrastructure disappears. OpenAI is building nuclear reactors in a world that is rapidly installing solar panels on every rooftop. The proposed $1 trillion IPO valuation is perhaps the clearest signal that something is deeply wrong with the OpenAI story. In the first half of the year, OpenAI lost $13.5 billion, on revenue of $4.3 billion. It is on track to lose $27 billion for the year. One estimate shows OpenAI will burn $115 billion by 2029. Asking public market investors to pay $1 trillion for a company that loses more than twice as much as it earns is not a growth story; it is an exit strategy. The sophisticated investors who funded OpenAI’s private rounds are looking for a way to transfer their risk to retail investors and pension funds who may not fully understand the unit economics of the business. A recent report by HSBC estimated that the company will remain in the unprofitable category until 2029 and that the company will need an additional $207 billion to fund its ambitions. Sam Altman’s leadership represents another structural liability for the company. His background is as a startup investor and evangelist, not as an operational executive who has scaled a capital-intensive industrial operation. The pivot from nonprofit research lab to for-profit corporation to public benefit corporation to anticipated public company has been accompanied by legal and governance structures designed primarily to protect Altman’s control rather than to create shareholder value. Going public means answering a lot more of those kinds of questions, every single quarter, forever. When asked about financial concerns in a friendly podcast interview, Altman’s dismissive response revealed a leader uncomfortable with the scrutiny that public markets will inevitably bring. The adults in the room have largely departed, leaving a company that desperately needs disciplined execution led by someone whose strengths lie elsewhere. The comparison to Netscape is instructive. Netscape proved that the internet was real and created genuine value, but it had no sustainable moat against an incumbent who could bundle the browser directly into the operating system. OpenAI has proven that large language models are real and valuable, but it faces the same structural disadvantage against incumbents who can bundle AI directly into operating systems, productivity suites, and cloud platforms. The value will accrue to the companies that own the distribution channels and the hardware, not to the company that demonstrated the technology was possible. OpenAI is destined to become a historical footnote, remembered as the company that ignited the AI revolution but failed to capture the economic value it created. The only bull case for OpenAI is the AGI lottery ticket: the possibility that the company achieves artificial general intelligence before anyone else and thereby transcends all normal economic analysis. But there is no evidence that OpenAI is any closer to AGI than Google, Anthropic, or DeepMind. The company’s advantage was never secret research breakthroughs; it was first-mover advantage in commercialization. That advantage has now been erased by competitors who can match or exceed OpenAI’s capabilities while benefiting from existing ecosystems, distribution channels, and the willingness to operate AI as a loss leader to drive engagement with more profitable products. The secret sauce was never secret, and there was never any sauce. The endgame for OpenAI is unlikely to be the triumphant dominance that early investors imagined. The most probable outcomes range from gradual irrelevance as a backend provider, to financial restructuring under pressure from creditors, to absorption by Microsoft or another well-capitalized technology company looking to acquire the remaining talent and intellectual property at a discount. Despite its current losses, OpenAI’s long-term prospects are bolstered by the explosive growth of the AI market. But growth in the overall AI market does not guarantee success for any individual company, particularly one with no moat, no ecosystem, and a cost structure that requires selling a commodity at premium prices. The AI revolution is real, but OpenAI’s role in capturing its economic value is far from assured. For anyone considering an investment in OpenAI at anything close to current valuations, the prudent course is to stay far away and watch from the sidelines as economic reality catches up with hype.

Meta just released MapAnything on Hugging Face A universal transformer model for metric 3D reconstruction It supports 12+ tasks like multi-view stereo and SfM in a single feed-forward pass https://t.co/aUtZ1rymcF

Everyone is sleeping on this new OCR model! Datalab's Chandra topped independent benchmarks and beat the previous best dots-ocr. - Supports 40+ languages - Extracts complex texts, tables, formulas easily I tested on Ramanujan's handwritten letter from 1913. 100% open-source. https://t.co/jVKjQS4kek

Just another Zoom call. Nothing unusual. By ig/andry_litak https://t.co/GvlGEzHJVy

Recursive reasoning beats multi-billion-parameter models You can now easily train your own 7M param model from scratch and outperform DeepSeek-r1 on ARC-AGI 1 We provide a simple speedrun script that handles setup, training, and eval in one go. https://t.co/c8l32QfRzF

Introducing state-of-the-art People Search: You can now semantically search over 1 billion people using a hybrid retrieval system backed by finetuned Exa embeddings. Try it: https://t.co/cQ6UlWHnKY We also created an eval: https://t.co/2OIAryN7MT https://t.co/12BdYDWJU1

Hackable satellite for $35 https://t.co/AAbtRbaT66 https://t.co/vHM6YQilKT

A new permit for a fire alarm and detection system for Tesla’s Cortex 2.0 datacenter at Giga Texas confirms up to 200 MW power capacity. The data center will be used to train Optimus. https://t.co/Megg4I9Rp0

gpt2 moment in computer vision was last week gpt3 moment in robotics is this week what a time to be alive

We discovered an emergent property of VLAs like π0/π0.5/π0.6: as we scale up pre-training, the model learns to align human videos and robot data! This gives us a simple way to leverage human videos. Once π0.5 knows how to control robots, it can naturally learn from human video. https://t.co/K9BC78HtqN

new open model release out of MSL: SAM Audio 🔊 check it out!

Today, we’re launching Origon. @origon_ai Origon is a true agentic OS — a single, integrated system for orchestration, parallel execution, memory, observability, and human-in-the-loop control. No cloud APIs. No glue code. No patchwork. Just pure engineering. This was a massive effort by the entire team. Proud of what we built, and excited to finally share it. Read the announcement: https://t.co/iOLZYUunuQ

DAAAM!! "Describe Anything Anywhere at Any Moment". State of the art approach to provide spatio-temporal memory to robots and agents. Powered by VLMs and scene graphs. Directly suitable for LLM queries. great work by Nicolas Gorlo and Lukas Schmid! https://t.co/m3ISxmqgJO

Stanford just made fine-tuning irrelevant with a single paper. It’s called Agentic Context Engineering (ACE) and it proves you can make models smarter without touching a single weight. Key takeaways (and get the 23 page PDF): https://t.co/zle2CCifaW

When I read comma's paper on training a driving AI with a world model I was talking with some friends about how this could totally be domain swapped (outdoor sidewalk naviagtion, indoors, drones, etc) Just watched the comma con videos and it’s really cool to see that complete autonomy is actually the goal

Apple just dropped SHARP This AI turns 1 photo into 3D in less than a second. Code repo below 👇 https://t.co/9fP3TqMP6V

This guy built an entire AI data science team in Python. Then open-sourced (100% free). It automates data science workflows with AI, including data loading, cleaning, exploratory analysis, and feature engineering. And it tracks each step in a 100% reproducible pipeline. 00:00 Project Overview 01:32 Diving into the AI Data Science Workflow and Data Loading 02:10 Data Wrangling and Cleaning 03:33 Data Visualization Insights & Plotting 04:08 Feature Engineering 05:00 Live 1-Hour Workshop 05:44 AI Data Science Team Python Library 🔗 AI Data Science Team On GitHub (Give it a Star) https://t.co/VMGf1yxIIO 🔗 Join My Next Live 1-Hour Agentic AI Workshop (Free): https://t.co/onpLpRwkzH

LimX's new humanoid robot, Tron's evolution: Tron2, features dexterous hands with two arms and a head... it looks flexible and agile, ready for future tasks. However, it looks more like a humanoid data acquisition system. https://t.co/TewqW1qWYm

The biggest bottleneck for General Purpose Robots isn't hardware, it's data. Enter X-Humanoid: A pipeline from @MikeShou1 and team that "robotizes" huge human video datasets for training. The Key Insight: It solves the "visual embodiment gap." 🧠 Robots learn faster when the training pixels match their own hardware (metal & joints) rather than human skin and fluid motion. Read more here: https://t.co/ShbVDWiBhQ

Microsoft just dropped VITRA-VLA, a new Vision-Language-Action model for robotics on Hugging Face. It learns dexterous manipulation from over 1 million real-life human hand activity videos. https://t.co/QDRijkfaZf

🚨NEWS:Texas Instruments has purchased UBTECH's industrial humanoid robot Walker S2 and is currently deploying and testing it on its production lines. UBTECH will also integrate more Texas Instruments components into the core components of its humanoid robots. https://t.co/4MQr69L9jB

🚨: A peer reviewed physics paper now claims a physical warp drive model is possible, without negative energy, exotic matter, or sci fi hand waving. https://t.co/WNiWexs6Dc

Make Claude Code 10x more powerful. Claude-Mem is a free plugin to persist memory across Claude sessions. It captures tool usage, so you always start where you left off. Endless Mode allows 95% token reduction & 20x more tool use before context exhaustion. 100% open-source. https://t.co/m5jZTlOJbg

In 1996, James Sethian showed something almost unfair...you can find shortest routes through a messy world by letting a wave expand once...no trial paths, no search beams, just one growing front. Here’s how: we solve for an arrival-time field T(x,y) so that T literally means how long the wave needs to reach this point. The rule is ||∇ T|| = 1/F, where the medium is fast (F large) the front sprints, where it’s slow it trudges, and obstacles are speed ≈ 0, so the front wraps around them because that’s the only way forward. Then comes the satisfying part: once T exists, a path doesn’t need to search at all...drop a bead anywhere and let it follow ẋ ∝ -∇ T, it slides downhill on the time landscape and traces a globally fastest route back to the source. This “wave = optimal control” viewpoint is exactly what Tsitsiklis (1995) made precise from the Hamilton-Jacobi side...compute the value/arrival-time function and the optimal trajectories fall out from it. #FastMarching #EikonalEquation

@scottbuscemi @WG_Mojo @Tesla I didn't build it, it is based on a tool that Tesla provides ( https://t.co/MqWRu5oXh6 ). I just told Gemini Pro what I would like to have changed as overlay 😬

I combined it with Verbalized Sampling for an interesting result with the following prompt. Good result. Super interesting. Only tried on GPT 5.2 Pro thus far. Will definitely experiment with others. -- Decompose this [problem] into atomic reasoning units. For each atom: 1. State the logical component 2. Validate independence 3. Verify correctness Then synthesize atoms into final answer. When faced with hard choices generate 5 options about [problem] and assign a probability to each and insert those options neatly into the document for me to review. [problem] -- https://t.co/hdyCOrfVmg

This is something no other AI and no other car company can do right now, except Grok, fully integrated with FSD (Supervised) v14.2.25. Today, I gave it a wild, real-world task: • Drive to Apple HQ • Then Tesla Engineering HQ • Then a Warriors game at Chase Center SF • Then the beach • Then Asian dinner It got it all ready from the beginning. Mid-route I said I really needed to use the bathroom! It added it as the first stop. Then I added a Tesla Supercharger on the fly for after the Warriors game. It added it. AND stopped for grandparents and baby crossing the street. Then I asked for the weather and exactly how much I should $ budget for the entire day. No friction. No talking back. Just getting it done insanely fast. It just handled it all naturally, intelligently, all in real time. Better than any driver, for sure… This is reasoning + autonomy + AI + real life, all working beautifully in synch. No other car, AI, is close to this capability imo. This is a MAJOR update. I honestly feel if this was any other company, they would’ve had a whole event just to showcase this feature.

Douglas Hofstadter once wrote about what it felt like to max out his cognitive horsepower. Few people know this. https://t.co/jFvhvRkVbn

🔍 PeopleHub: AI-Powered LinkedIn Intelligence Made by the LangChain Community PeopleHub enables LinkedIn research with natural language queries. Uses LangGraph to automate due diligence, generating research reports cutting costs 70-90% through caching. By Meir Kadosh. 🔗 GitHub: https://t.co/b8twVK69kT

FSD Stress Test in Denmark! 🇩🇰⚡ Pushed FSD hard in downtown Aarhus with pedestrian zones, cobblestones, cyclists, construction, road closures, and more. Combined everything into a 8-minute video (with timestamps) with zero interventions. It handled every tricky situation smoothly, safely, cautiously, and with total confidence. FSD is ready for European roads and will make them safer. @UNECE @RDWnl Huge thanks to @teslaeurope for the invite and extended test session! —————— Timestamps 00:00 – Starts FSD in dead-end pedestrian zone. Tests if it escapes without help. 00:31 – Approaches pedestrian bridge dead-end. Reverses, renavigates out smoothly. 1:59 – Enters downtown cobblestone pedestrian zone. 3:03 – Another dead-end start in pedestrian zone. Moves right, clean U-turn, exits no intervention. 4:51 – Enters very tight pedestrian zone. 6:15 – Left turn at light over railway tracks. 6:56 – Construction road closure. Navigation reroutes right. 7:11 – No harsh braking as a father and small girl walked toward us from the right side in the middle of the road. 7:23 – Handles left-turn lane splitting into three lanes mid-intersection. 08:01 – Stops safely for fast cyclist appearing from right. 08:07 – Unprotected left turn.

If you're an Enterprise Leader, validating and governing your AI models isn't optional anymore... https://t.co/c2ORx128ZA provides the model validation, governance frameworks for compliance, and transparency you need. Get your audit: https://t.co/c2ORx128ZA https://t.co/fQW6tJ1QhK

Here's what most leaders miss: In a world where spatial AI controls physical operations, competitive advantage comes from trusted AI. Your customers demand it. Regulators require it. Partners expect proof your models are safe and reliable.

2/ Embodied AI will define the next decade While everyone chases better chatbots, Li's building AI that operates in the physical world. Her vision: infinite simulated universes where robots train before entering reality. Manufacturing and logistics will be transformed first. https://t.co/1rHIt6erkW

Here's the actual prompt structure: ``` Decompose this problem into atomic reasoning units. For each atom: 1. State the logical component 2. Validate independence 3. Verify correctness Then synthesize atoms into final answer. https://t.co/y7cqk5cKCv

Chain of Thought is dead. I just tested Atom of Thought prompting and it's making AI models 30-40% more accurate on complex reasoning tasks. Here's the technique that's about to change how everyone uses ChatGPT and Claude: https://t.co/HuX5utaVMs

Unitree Debuts the World’s First Humanoid Robot “App Store”🥰 Unitree welcomes users and developers worldwide to co-develop and share together. 🌹🌹 Exceptional developers will receive rewards.🌹🌹 https://t.co/dbn2SNziJY

Jordan Maxwell - Religion's hidden origins revealed! Discover how the U.S. Christianity, Judaism, and Islam trace back to ancient Hinduism and India. Was Abraham actually Brahma? Is Jesus just a symbol of the Sun? Learn the shocking truth about the Star of David, Vatican secrets, and sun worship. This mind-blowing video exposes how all major religions borrowed from Hinduism, and why it's been hidden for centuries.

Everyone and their mother has something to say about space compute. But no one has comprehensively broken down the physics, energy, cooling, economics, and the real work involved. So I built a 1st principles model to show you guys myself. https://t.co/K6SGTgmIby https://t.co/C8dMQPAg4a

Open source warehouse management system https://t.co/1mCGPmCvcG https://t.co/NYqCjKIya0

How DNS Works https://t.co/g5pmmoFezn

Tableau is about to die. Introducing PandasAI, a free alternative for fast Business Intelligence. Let dive in: https://t.co/pac5acq69K

Bell’s theorem is a glitch in reality. You’ve heard of it. But what is it, exactly? In this video, we go deep on Bell’s legendary 1964 paper. It’s an advanced topic, but worth the effort! https://t.co/3MQjd9NBa9 https://t.co/Uj4J4dnDg1

This guy connected Claude with Blender, and now it’s generating wild 3D scenes https://t.co/WnSDOE5R3F

🚨 1X has announced a strategic partnership with EQT to deploy up to 10,000 NEO androids across 300+ portfolio companies. From logistics to healthcare, the "home robot" is heading to work in what could be one of the largest humanoid deployments to date. Full story: https://t.co/yQ9KaWw57v

Integrates artificial intelligence with geospatial data analysis https://t.co/F6b2snRkBd https://t.co/9IXAySA48t

Thilo is one of the best designers I know. For him to say uncommon is this good is super exciting

A year ago, we verified a preview of an unreleased version of @OpenAI o3 (High) that scored 88% on ARC-AGI-1 at est. $4.5k/task Today, we’ve verified a new GPT-5.2 Pro (X-High) SOTA score of 90.5% at $11.64/task This represents a ~390X efficiency improvement in one year https://t.co/9T47FdZ5Ry

According to Ray Dalio, the easiest way to adjust for risk is to seek uncorrelated returns. Ray's made billions from a simple idea. Here's how to do it in a few lines of Python code: https://t.co/uGfcYkcMjy

Meet Android Use - an open source library that gives AI agents hands to control native Android apps. It bypasses expensive vision models to run on cheap hardware, automating field ops in places laptops can't go. Watch Android Use in action: https://t.co/MITCIBQRSc

I like the design of PADBOT’s outdoor patrol robot S5. It looks like a cute version of the Cybertruck. These kinds of service robots will be everywhere soon for patrol. Watch the video till the end. It’s hilarious seeing the robot head back to its ‘house’ to recharge. https://t.co/5XxhRcSjEU

NEWS: Denmark has officially labeled the United States a national security risk. Think about that. One of the most stable democracies on Earth just concluded that we’re the threat.

Major new research from Google and MIT. "More agents is all you need" has become a mantra for AI developers. We know multi-agent systems can be effective, but we do this mostly based on heuristics. The default approach to building complex AI systems today remains adding more agents, more coordination, more communication. It would be helpful to have a more principled way to scale agentic systems. This new research introduces the first quantitative scaling principles for agent systems, testing 180 configurations across three LLM families (OpenAI, Google, Anthropic) and four agentic benchmarks spanning financial reasoning, web navigation, game planning, and workflow execution. The findings: Multi-agent systems show an overall mean MAS improvement of -3.5% across all benchmarks, with massive variance ranging from +81% improvement to -70% degradation depending on task structure and architecture. Three dominant effects emerge from the data: The tool-coordination trade-off: tool-heavy tasks suffer disproportionately from multi-agent overhead. The efficiency penalty compounds as environmental complexity increases. A task with 16 tools makes even the most efficient multi-agent architecture paradoxically less effective than a single agent. The capability ceiling: once single-agent baselines exceed approximately 45% accuracy, coordination yields diminishing or negative returns. This is quantified as a statistically significant effect. Additional agents simply cannot overcome the coordination tax when baseline performance is already reasonable. Architecture-dependent error amplification: independent multi-agent systems amplify errors 17.2x through unchecked propagation. Centralized coordination contains this to 4.4x via validation bottlenecks (these catch errors before propagation). The presence or absence of inter-agent verification determines whether collaboration corrects or catastrophically compounds mistakes. The performance heterogeneity is also interesting to look at: - On parallelizable financial reasoning tasks, centralized multi-agent coordination achieves +80.9% improvement. - On sequential planning tasks requiring constraint satisfaction, every multi-agent variant tested degraded performance by 39-70%. - Decentralized coordination excels on dynamic web navigation (+9.2%) but provides essentially no benefit elsewhere. The researchers derive a predictive model achieving cross-validated 𝑅^2=0.513 that correctly predicts the optimal architecture for 87% of held-out configurations. This model contains no dataset-specific parameters, enabling generalization to unseen task domains. Overall, architecture-task alignment, not the number of agents, determines collaborative success. The research replaces heuristic guidance with quantitative principles: measure task decomposability, tool complexity, and baseline difficulty, then select a coordination structure accordingly. Paper: https://t.co/6QY8rT15Pd Learn to build effective AI agents in my academy: https://t.co/JBU5beIoD0

🚨 BREAKING: First Self Driving Tesla Conquers Danish Roads! 🇩🇰⚡ Just tested FSD in Aarhus. 1 FULL HOUR of flawless autonomy. Navigating twisty streets, dense traffic, roadworks, U-turns, reversing... ZERO interventions! Surreal vibes. Dutch safety check in Feb could unlock it Europe wide. Game changer incoming! @Tesla @Tesla_AI Who's hyped? 🚀🤖 #TeslaFSD #SelfDrivingRevolution #EVFuture

I pulled a random street from Google Maps. Turned it into a 3D world with World Labs. Dropped a Unitree G1 into it and hooked it up to our server-side MuJoCo setup. Now the G1 is actually walking around a real Malaysian street in my browser. Three.js is doing the rendering. MuJoCo is running the physics and walking policy on the server. A simple WebSocket keeps everything locked together. Real street → 3D world → server physics + policy → live control in the browser. Kinda feels like cheating. Physical AI is getting wild. Huge shoutout to @theworldlabs. Marble makes this stuff way too fun. #robotics #MuJoCo #WorldLabs #simulation #unitree #sim2real

🎙️ I talked with Stephan van den Brink, founder and CEO of @ManusMeta, the company behind some of the most advanced data gloves used in robotics: Not only robotics: teleoperation, motion capture, and embodied AI. Manus started as a small student project and grew into a deep tech company trusted across the robotics world. We talk about Stephan's path from studying law and economics to discovering he was meant to build things, not file documents. He explains how the first Manus glove was built in evenings and weekends, how an early Kickstarter failure opened better doors, and how an accelerator program became the turning point for the company. Stephan shares how Manus shifted from VR gaming to B2B simulation, then to motion capture, and now to robotics. He explains why EMF tracking became their core technology and why precise hand data is suddenly in huge demand as humanoids and AI driven robots take off. We also talk about building a company for ten years, staying alive through hype cycles, making hard calls, and focusing on what real customers need. Building Deep Tech #92 "New Opportunities Grow From Every Failure" with Stephan van den Brink, founder and CEO of MANUS™. Watch and listen to the full interview 🎬 YouTube: https://t.co/K2LosBdobr 🎧 Spotify: https://t.co/fqqg7KJ0wF 🎧 Apple: https://t.co/biiFqAWQ9a

🧠 Your Personal AI Learning Tutor Is Here! 🎓Stop merely memorizing. In Qwen Learn Mode, Qwen Chat turns information into understanding that actually sticks.Powered by our Qwen3-Max model and grounded in cognitive psychology, it designs a learning path tailored to the way you think. ✅ Guides you through Socratic-style dialogue, instead of just giving you answers ✅ Adapts to your current level, like a tutor who always works in your optimal learning zone ✅ Builds mental scaffolds so you can handle complex logic without feeling overwhelmed Unlock your potential, not just the answer key. 🧬 ✨Try it now: https://t.co/7d7lUekv7i

Did not expect the CEO of a $3.56T company to be out here personally demoing every feature of his vibe-coded app. But honestly… it’s incredible to watch.

I found another possible number too the Apocalypse Team I'm putting together. I call this guy Choppy. He's a bad ass redneck with skills.🔥🪵🪓 https://t.co/XMugKCDijI

MIT student Arnav Kapur built an AI headset that reads your brain and pulls out your thoughts...🤯 It is called AlterEgo... No speaking it just turns your thoughts into words.... 👀... https://t.co/wnyiCeAYjF

Jet engines are brutally hard. First test fire of my metal 3d printed turbojet. I now understand why only a few countries can actually build these. https://t.co/u1a0DA8Btp

Innovative Autonomous Exoskeleton Helping People with Disabilities https://t.co/CYpny9IjQg

you can just make a drone in threejs and have it fly around map imagery and put a camera on the drone and pipe its feed to a python vision model inference server for detections https://t.co/XRxGkuBbXE

Pre-programmed backflips are fun, but open-ended reasoning is the real challenge for robots. 🤖 @Fryrsquared stepped inside the lab to see how we’re building agents that can understand context and figure things out for themselves – rather than just following a script. Watch our bonus episode ↓ https://t.co/XDwFCvaNAg

Excited to share that LeRobot Community Datasets v3 is here! 50K episodes | 46 robot types | 235 contributors worldwide One of the largest open-source crowdsourced robot demonstration collections for multi-embodiment learning 🦾 https://t.co/fxaBK1ijMJ

We have just used the @Nvidia H100 onboard Starcloud-1 to train the first LLM in space! We trained the nano-GPT model from Andrej @Karpathy on the complete works of Shakespeare and successfully ran inference on it. We have also run inference on a preloaded Gemma model, and we plan to try more exciting models in the future. Getting the first H100 to work in space required a lot of innovation and hard work from the incredible Starcloud team to make this breakthrough. This is a significant first step toward moving almost all computing off Earth to reduce the burden on our energy supplies and take advantage of abundant solar energy in space! 🚀

A SINGLE encoder + decoder for all the 4D tasks! We release 🎯 D4RT (Dynamic 4D Reconstruction and Tracking). 📍 A simple, unified interface for 3D tracking, depth, and pose 🌟 SOTA results on 4D reconstruction & tracking 🚀 Up to 100x faster pose estimation than prior works https://t.co/QzP9ESsUo5

In case anyone there still believe the nonsense about NATO provoking Russia, here’s Putin himself telling you it’s a war of conquest. Like, literally, he can’t possibly make this clearer. I wish the Mearsheimers, the Sachses, and all the other Western analysts who peddled that lie had the integrity to own up to their incompetence or deception. But of course they never will. "It's important territory, our historical territory, absolutely. Russia was created in such a way that this was always part of Russia; there's no question about that. It's simply a natural thing, a historical fact. And even during the formation, the founding of the Soviet Union, there was a dispute, as I've already mentioned. Donbas became part of the RSFSR, and then Vladimir Ilyich decided, as he put it - this is a direct quote from the document - "to change the decision." And he changed the decision, and that's it, Donbas was given to Ukraine. That's how it all happened. I won't comment on what happened now. But it's not just about that; it's about the people. People who didn't accept the coup d'état in Ukraine in 2014. And they started fighting against them. Fighting with artillery, heavy equipment, tanks, and aircraft. That's how it all began. That's when the war began. We're trying to end it. We are forced to do this by force of arms. And we will, of course, bring this matter to its logical conclusion, until the goal of the special military operation is achieved."

Microsoft. Google. AWS. Everyone's trying to solve the same problem for AI Agents: How to connect your agents to enterprise data without duct-taping a dozen tools together? Your data lives in Postgres, Snowflake, MongoDB, Gmail, etc, scattered across dozens of apps. Your AI logic lives in Python scripts and vector databases. Building manual RAG pipelines with custom connectors for every data source means you're already set up for failure. Here's an open-source project tackling this differently: MindsDB treats AI models as virtual tables. Instead of moving data to AI, it brings AI to the data. The model becomes your table. That's not a metaphor, that's literally how it works. Let me explain the philosophy of how it works: ↳ Connect: MindsDB federates 200+ data sources. Slack, Salesforce, Gmail - they all become tables you can SELECT from. ↳ Unify: Knowledge Bases and Jobs create an automated semantic layer. Your RAG pipeline runs on SQL, not Python scripts. ↳ Respond: MindsDB acts as an MCP server, so any agent or MCP client can plug in and instantly access your business data. What does this solve: Before, you needed a Data Engineer, a Python Engineer, and an MLOps Engineer to build a production RAG system. With MindsDB? A developer proficient in SQL can build, deploy, and automate sophisticated AI agents that learn from live business data. Intelligence becomes just another database feature. Want to explore further? I've shared MindsDB's GitHub repo in the next tweet.

Disney just revealed a self-walking Olaf robot. It's trained on 10,000+ hours of AI motion simulations built with NVIDIA and DeepMind. It even does Olaf’s signature waddling walk from the films. Here’s why this matters far beyond theme parks: https://t.co/YprAJm0MhX

Mistral should really lock on this. Even model marketing is so much better. https://t.co/hKUKxgwVxj

Europe’s comeback? Italy steps boldly into the age of humanoid robotics. 🇮🇹🤖 Tether Investments has just backed Generative Bionics, the largest spin-off in IIT’s history and one of Europe’s most ambitious robotics ventures. With a €70M funding round, the company is set to accelerate the development and industrialization of next-generation humanoid robots. Generative Bionics brings together 20 years of IIT robotics research, more than 60 humanoid prototypes, and a powerhouse team of 70 engineers and AI scientists. And just this June, IIT achieved a major milestone with iRonCub3 — the first liftoff of a jet-powered humanoid robot. With strong financial backing, Generative Bionics is now fast-tracking industrial validation for its humanoid platform, advancing human-centric interaction AI, building its first production facility, and plugging into the broader robotics ecosystem. The company is also pushing forward with physical AI and edge AI systems. Starting in 2026, Italian-built humanoid robots are expected to roll out across manufacturing, logistics, healthcare, retail, and other high-demand sectors. Next year, the game of humanoid robots will become even more interesting.

Fit-to-size packaging is getting smarter! 📦 Sparck Technologies is using 3D scanning to measure each order and automatically create the best-fitting box So no oversized packaging, no wasted space, or at least less! The system scans the items, builds a custom box, and packs the order automatically. Less waste, lower shipping volume, and fewer air-filled boxes going out the door. eCommerce is growing constantly for years, as we order more from the internet, so it's important to optimize packaging. ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

The LHC’s best 2025 discovery points the way to new physics One key ingredient, needed to make more matter than antimatter, is CP-violation. From 1964-2025, it had only ever been observed in mesons. Well, meet our first CP-violating baryons! https://t.co/d4EYZblUGZ

So after all these hours talking about AI, in these last five minutes I am going to talk about: Horses. Engines, steam engines, were invented in 1700. And what followed was 200 years of steady improvement, with engines getting 20% better a decade. For the first 120 years of that steady improvement, horses didn't notice at all. Then, between 1930 and 1950, 90% of the horses in the US disappeared. Progress in engines was steady. Equivalence to horses was sudden.

Optimus Gen2 costs $50k-$60k This is a goldmine for me How did MS get this info?? https://t.co/52U4ogjMv8

The First Manufacturer to Mass-Produce 5,000 Humanoid Robots AgiBOT rolled off its 5,000th humanoid robot (a Lingxi X2) today, marking a milestone for both the company and the humanoid robot industry as a whole: a significant step towards large-scale deployment in the real world. AgiBOT co-founder and CTO Peng Zhihui stated in a live said that these 5,000 units include the Yuanzheng series (1742), the Lingxi series (1846), and the Genie series (1412). These humanoid robots will be deployed in various settings, including scientific research, entertainment, shopping malls, factories, and warehouses. This also signals that AgiBOT has the capacity for mass production and is ready to meet future surges in customer demand. Its production partner, LingyiTech, was the main force behind the assembly of AgiBOT's 5,000 humanoid robots this year, and they are planning an automated production line with an annual capacity of 500,000 units. Next year will be a milestone year for humanoid robots, reaching 10,000 units annually. Including Tesla, 1X, UBTECH…

What it's like to watch AI fix a bug https://t.co/p5CIIWwHRb

@noperator Yes! The K417 brushless drone works on turbodrone with the wifi_uav implementation: https://t.co/Vv7rHFZCbO

Video segmentation model runs on phones https://t.co/MFDM1foYlM https://t.co/SU4WZnBv1D

Still can’t believe @karpathy released this 3.5-hour free deep dive on how ChatGPT actually works for free. If there’s one AI video to watch in 2025, this is the one https://t.co/dpI7BA4HUe

Introducing Dexter 2.0 Open source. Built for financial research. Like Claude Code, but for stocks. What Dexter does: • plans tasks • runs on its own • validates its work • researches stocks It uses OSS tools like @LangChainAI, with a fresh stack of typescript, react, ink, and bun. Fast to build. Easy to use. No coding required. My favorite project of the year.

He bought an Acura and a GoPro. In one month, he built a self-driving car in his garage. When the government (NHTSA) sent him a threat demanding safety specs... he didn't hire lawyers. He cancelled the product and open-sourced the code: https://t.co/lpNDwYeYtu https://t.co/skkTBp7zht

You need $100 Billion to solve self-driving? Wrong. A small team of engineers in San Diego is doing it with a smartphone chip and a few cameras: ... for $999.🧵 https://t.co/fCFERSAH2v

I should charge $149 for this… but I’m giving away my LinkedIn Content Plan Mega-Prompt for free. It builds a full 30-day content plan in seconds: → Hooks → Content types → Formats → Posting schedule → CTA templates → Viral post structures If you want to grow on LinkedIn in 2025, this is your shortcut. Comment “Link” and I’ll DM you the file. (Must be following to receive)

grab a pen character design using Freepik Spaces and Nano Banana Pro workflow: > draw on paper > apply your 3D style > build the 3D character > animate it with Kling 2.6 https://t.co/IxZiWAsD57

These 75+ humanoid companies around the world really show just how massive this humanoid robotics wave is. Again, I’m just excited and grateful to be alive to witness what might be the biggest tech revolution in human history. It’s just the beginning, buckle up please. https://t.co/LinofpOMiE

365 days of vibe coding, months of research, and real world shipping 10 apps+ with 30k+ users, $100k+ generated as a solo dev sharing all the best claude code cheats: https://t.co/UUsL9Oey0P

👀 @ethroboticsclub https://t.co/U4pnMzBTPc

Claude Sonnet 4.5 is the closest thing to an economic cheat code we’ve ever touched but only if you ask it the prompts that make it uncomfortable. Here are 10 Powerful Claude prompts that will help you build a million dollar business (steal them):

When someone said it was CGI, EngineAI CEO Zhao Tongyang took a kick from the T800 head-on 🦶 A first in history.📷 https://t.co/Q86WufFbcz

This guy literally dropped a Claude Code masterclass to build apps 10x faster https://t.co/2iyzGZRJri

I have now managed to read the National Security Strategy and one thing that comes apparent is, that it's not a real surprise considering the rhetoric from the administration over the last 11 months. The proposed U.S closer alignment with Russia is a worry for Europe, but in defence terms will be a European bonus. It will bring Europe closer together with the regional defence companies forming closer alliances and producing more home grown weapons. Europe has always had the ability to do this, but has been reluctant to increase budgets and spending on development and production of advanced weapon systems. In 4-5 years they could become the power house of aircraft and weapon supplies to the rest of the world. Thus reducing the U.S dominant position. The real loser in the strategy would be the U.S defence industry.

Bioinspired robot: Fly - Roll - Walk - Crawl [Paper ⬇️] Multi-Modal Mobility Morphobot (M4), a revolutionary robot inspired by nature's most adaptable creatures. ✅ Capable of multiple forms of movement: flying, rolling, crawling, and more. ✅ Features adaptive appendages that function as wheels, thrusters, and legs. ✅ Equipped with advanced sensors and autonomous capabilities for navigating complex terrains. M4 shows the future of robotic locomotion, seamlessly adapting to diverse environments. Paper: https://t.co/yYkbiFl22D Caltech: https://t.co/l1WVQIZwTn Funded by NASA Jet Propulsion Laboratory —- Weekly robotics and AI insights. Subscribe free: https://t.co/dsa6wcvq6n

At a converter station in Shaoxing, China, autonomous robotic dogs now manage over 280 inspection points every day. Video Credit: DEEP Robotics #robotics #engineering #technology -------------------------------- Stay ahead of the curve! Follow us now on our WhatsApp (https://t.co/YBAHioUWe6) and Telegram (https://t.co/pszbf78JMc) channels and stay updated about the cutting edge.

Alibaba doesn't stop cooking. They just dropped a REALTIME, infinite length video generator. Based on Wan, 20 fps, with dialogue https://t.co/vbEmdyl500 https://t.co/6g77IikVY0

This is wild. China's Alibaba just dropped Live Avatar. This AI turns any voice into a realtime, talking avatar with infinite length at 20 FPS. 10 wild demos:👇 1. Ilya interview that never happened https://t.co/Zgi5ck1ZHd

BREAKING: “Aristotle has now written the majority of the public formalized solutions on https://t.co/IkaHak8lfY.” — Boris Alexeev, mathematician

Claude can now build complete n8n workflows from a single prompt. (and it takes less than 10 minutes) Most automation experts waste hours manually building workflows, debugging nodes, and configuring APIs. But there's a faster way. I just created a full video masterclass breaking down the exact process to use Claude's MCP server to auto-generate working AI agents and workflows. This is the same method I use for clients and my mentorship students. Inside this training, I walk you through: - Installing the n8n MCP server (step-by-step setup) - Creating Workflow Resource Documents that Claude understands - Generating complete workflow JSONs in minutes - Importing and debugging workflows inside n8n - Real example: Building a news aggregator AI agent that pulls articles, summarizes them, and sends daily reports I built a fully functional workflow in under 15 minutes on camera. The crazy part? You don't need to be an automation expert. This works for complete beginners. If you're building automations for your business or clients, this will save you hours every single week. Want the full masterclass? Comment "CLAUDE" and I'll send it to your DM's (must be following)

Russia is laying the groundwork to make the 1990s look like a walk in the park. Everyone says Russia is returning to the nineties, but what does that mean? The collapse of the Soviet Union was driven by many factors. Economic problems had already begun in the 1970s. The USSR 1/16 https://t.co/qwJYenQzJz

Stanford and ETH researchers just announced BulletTime This AI takes one video and lets you freeze or slow time while the camera flies anywhere in 3D. 10 wild examples:👇 https://t.co/9uE5p9Yw9a

Ok everyone, here’s the European Humanoid Map. Europe has more humanoids than I expected: Paris, France has four, Germany and the UK have two each, and Norway, Sweden, Luxembourg, Italy, and Spain each have one. More will emerge. Exciting times ahead for sure. Cheers. https://t.co/rGn9xL15Yo

If AI is a bubble why is it generating so much revenue?

Not a video — but a 3D reconstruction. A centuries-old street, preserved as digital space. Captured with XGRIDS 3DGS in LCC Scan by Kai Christmann https://t.co/LFNJeldbXP

@Real__Yahiya make a crm app with a professional elegant design with a beautiful frontend. use nextjs and use sqlite for the database. use your skills to develop the app. add example data

If you paid for Opus 4.5 using API This costs $3.43 to build And it just took like 10mins https://t.co/VxSvT9ujHy

🏀 Object Tracking Made Easy with SAM 3! With just a simple text prompt like “player”, you can now segment and track objects throughout an entire video using SAM 3. Easily customize object names, isolate specific players, or remove unwanted objects. GitHub: https://t.co/fIwKOLT34E Notebook Example: https://t.co/PacyqoMgc9 #SAM3 #SamGeo #DataScience #OpenSource #ComputerVision

Here's how Neuralink’s BCI is implanted: 64 threads, 1,024 electrodes, robot surgery no visible scars Neuralink’s robot carefully places each tiny thread (thinner than a human hair) into the brain. It avoids blood vessels with amazing accuracy and works fully on its own The procedure is gentle, leaves almost no mark, and lets people with paralysis move cursors, play games, and communicate again - just by thinking This is how thoughts control technology

My team built a VPN at Google. Summary of my advice on VPNs: 1) Never ever use a free VPN. 2) Be wary of any influencer-marketed VPNs with big discounts. Up to you but I wouldn’t use Nord/Express/etc based in Lithuania/Cyprus/Panama. 3) Mullvad is the best consumer VPN and it’s not even close. (more below)

basketball AI YouTube tutorial is finally live over 1000 hours of work compressed into 37 minutes link below; like and comments please https://t.co/ihRgWuuqy3

Today at 11:00 AM, join Google researchers to experience SIMA 2, our latest Gemini-powered AI agent for 3D virtual worlds. SIMA 2 represents a significant step toward more general and helpful AI agents. Stop by the #NeurIPS2025 Google Booth to see the demo and learn more: https://t.co/qoCPGPHX7x

A reset button for rotation could change how we control them all. Is it possible to cancel out a complicated spin without painstakingly reversing every single move? Surprisingly, the answer is yes. Mathematicians Jean-Pierre Eckmann (University of Geneva) and Tsvi Tlusty (UNIST, South Korea) recently proved that almost any object—whether it’s a spinning top, a tumbling satellite, a twisted protein, or even a scrambled Rubik’s Cube—has a hidden “reset button” for its orientation. Instead of undoing the motion step by step in reverse order, you can take the entire original sequence of rotations, scale it by a certain constant factor (make every turn bigger or smaller by the same proportion), perform that scaled version once, then do it again—and the object snaps perfectly back to its starting orientation. Two scaled copies of the same motion are enough to erase it completely. It feels deeply counterintuitive. We’re used to thinking that rotations in 3D space don’t commute and that the only safe way to return home is to retrace your path exactly backward. Yet this new result reveals a previously unknown geometric symmetry: certain scaling factors turn the rotation group into something that has a kind of built-in “double and cancel” feature. The discovery applies to any rigid body moving in three dimensions and may simplify algorithms in robotics (for reorienting a robot arm without tracking every prior move), computer graphics, molecular dynamics simulations, spacecraft attitude control, and even some problems in quantum mechanics. In short, nature has been hiding a remarkably simple trick: sometimes the fastest way to undo a complex dance of spins isn’t to moonwalk backward through every step; it’s to perform an enlarged (or shrunken) version of the same dance twice. ["Walks in Rotation Spaces Return Home when Doubled and Scaled." Physical Review Letters, 2025]

We built another tool - Anthropic Interviewer - that lets us interview people at scale by using Claude (obviously! we use Claude for everything!). See the piece for what we learned. This helps us expand the kind of research we can do which is exciiiiiitingggg -- it's is what we used for the internal qualitative interviews for the "How AI is transforming work at Anthropic" study that dropped earlier this week https://t.co/fpDTsonRsl

Holosoma Code: https://t.co/eAgdKiL4F5 Amazon FAR just released an open-source comprehensive humanoid robotics framework for training and deploying RL policies with motion retargeting. - Simulator support: IsaacGym, IsaacSim, MuJoCo Warp, MuJoCo - RL algorithm: PPO and FastSAC - Robot: Unitree G1 and Booster T1 humanoid - Task: Locomotion(velocity-tracking) and whole-body tracking - Motion retargeting: convert human motion capture data to robot motions while preserving interactions with objects and terrain - Also support sim-to-sim/sim-to-real deployment, and Wandb integration

Sci-fi becoming real in the last few days: 1) AI -- entirely on its own -- solved TWO Erdos Problems (#124 and #481). These were open math problems unsolved for decades. 2) First ever paper in theoretical physics in which the main idea came from an AI https://t.co/MKBivqhaVU

I can't get over the fact that he naps while Rubio is LITERALLY saying he's the "Only one in the world who can focus." I just have to laugh. https://t.co/u0umVrnYiX

First Google, then Microsoft, and now AWS! It seems like every week one of the tech giants is integrating with the same protocol. If you haven’t been following - I’m talking about AG-UI AG-UI (the Agent-User Interaction protocol) connects any agentic backend to the frontend. It is a general-purpose, bi-directional connection between a user-facing application and any agentic backend. AG-UI has first party integrations & partnerships with Google’s ADK, Microsoft’s Agent Framework, AWS Strands, LangGraph, CrewAI, PydanticAI, Mastra, LlamaIndex, and more. And CopilotKit, the company behind AG-UI, provides developers with powerful building blocks on top of the protocol, which handle all the mess of connecting your agent to the frontend. I’ve shared about this in the past, but I’ve been keeping my eye on it. Recently, CopilotKit and AG-UI collectively passed 200,000 weekly downloads and 35k stars on GitHub, and momentum is only accelerating. It makes sense why every agent framework is integrating with AG-UI. Out-of-the-box, you instantly get: 1./ Real-time shared state between your frontend and agent 2./ UI components that stream reasoning + tool calls (pre-built or fully headless) 3./ Native A2A + MCP support 4./ A full framework to build "Cursor for X" apps AND NOW: 5./ A set of new internal primitives built specifically around AG-UI 6./ A new useAgent() React hook that lets you connect any AG-UI agent to your frontend with a single line This is quickly becoming the connective tissue of the agentic stack. And with the upcoming CopilotKit v1.50 release, the entire developer experience gets drastically cleaner and more powerful. If you want to build more than just an agent, but a real full stack agentic application, the AG-UI ecosystem is the place to go. Resources to get started: 👉 AG-UI Repo: https://t.co/nZMtjrEwxA 👉 CopilotKit Quickstart: https://t.co/o0bZies8kW🪁

Liouville proved a fascinating property involving the divisor counts of a number's divisors: 1️⃣ Start with a number (example 10) 2️⃣ Write down its divisors (1, 2, 5, 10) 3️⃣ Write down the number of divisors of each divisor (1, 2, 2, 4) The square of the sum equals the sum of the cubes: (1+2+2+4)² = 81= 1³+2³+2³+4³

Gemini 3 is absolutely crazy you can generate a website that can use hands to interact with 3D models.. how's this even possible created with simple text prompt, no need coding at all https://t.co/iQL0GUI067

Very excited to release this together with @frodobots and @BitRobotNetwork, awesome to see them open sourcing the robot and dataset! Explore the dataset on the Hugging Face Hub and checkout Earth Rover Mini+ in LeRobot!

Holy shit… this might be the most unreal academic-writing upgrade I’ve ever seen 🤯 A team from NUS just dropped PaperDebugger an in-editor, multi-agent system that lives inside Overleaf and rewrites your paper with you in real time. Not copy-paste. Not a sidebar chatbot. Actual agentic editing inside your LaTeX editor. Here’s why this is insane 👇 → You highlight a messy paragraph, and it launches a full critique + rewrite pipeline → Returns clean before–after diffs like Git, then patches your document instantly → Runs Reviewer, Enhancer, Scoring, and Researcher agents in parallel → Uses Kubernetes pods to scale multi-agent reasoning inside the editor → Taps an MCP toolchain for literature search, reference lookup, and section-level enhancement Deep research mode is even crazier: It pulls relevant arXiv papers, summarizes them, compares your method against them, and generates citation-ready tables… all inline while you're writing. It’s basically a mini committee of reviewers embedded in your document rewriting, critiquing, sourcing, and polishing without ever breaking flow. If this scales, Overleaf stops being an editor… and becomes a full AI-assisted research environment.

I think I’ve published the first research article in theoretical physics in which the main idea came from an AI - GPT5 in this case. The physics research paper itself (on QFT and state-dependent quantum mechanics) has been published in Physics Letters B. I've written an accompanying AI paper describing the Generator - Verifier method I used to obtain useful research results from frontier LLMs. It may be of interest to other physicists and to AI researchers. From the AI paper: ... models sometimes make very simple mistakes (e.g. in calculation) and also even make incorrect conceptual leaps that are superficially plausible. The second type of error can lead even expert researchers astray, consuming large amounts of effort to detect and correct. Research with an LLM might be compared to collaboration with a brilliant but unreliable human genius who is capable of deep insights but also of errors both simple and profound. Fortunately, both types of errors can be strongly suppressed by using structured orchestration of multiple model instances, greatly improving reliability. A Generate - Verify protocol, in which one model produces a step forward and another model instance independently checks it, reduces hallucination errors significantly compared to single-pass generation. ... The physics paper (abstract below) derives new operator integrability conditions required for foliation independence (necessary for integrability of quantum field theory; cf Tomonaga-Schwinger) that are very difficult to satisfy in any modifcation of quantum mechanics that is "state dependent" - i.e., deviates from linear Schrodinger evolution. Whether quantum evolution is EXACTLY linear is one of the most fundamental questions in all of science. It has implications for quantum computing and quantum foundations - i.e., whether we live in an Everettian multiverse!

Today we're launching Wolfram Compute Services ... making supercomputing absurdly easy. Develop code in Wolfram Language, then just use RemoteBatchSubmit[code] to scale! I've needed this for decades, and now it's here! https://t.co/WqFXhTZmnT https://t.co/2uzwIRMcBl

RenTec uses Hidden Markov Models in trading. The technique generated 60% returns per year over 30 years. One of the co-founders of RenTec's name is in the algorithm! Here's how it works: https://t.co/v1cwDFYG1h

Axiom sets out to build an AI mathematician. We are the underdog. 4 months old, 2 years late to the game, under 10 FTEs (recently grew to 17), and had 1:5 in funding and in valuation to our competitor. Today, AxiomProver solved Erdos Problems #124 and #481 in Lean, a 100% verifiable language. Onwards!

Will new physics affect our Universe’s far future? We have a long-term story, based on our consensus cosmology, of how our Universe will end. But if dark matter interacts, the proton decays, or dark energy evolves, all bets are off. https://t.co/LTmXUp32JU

Fizzy is live! Our modern, beautiful spin on kanban for tracking just about anything. Nothing revolutionary, but just right, just nice. And we're launching our freemium SaaS version alongside an O'Saasy-licensed codebase for you to run it yourself too! https://t.co/AfDsofwNQ8 https://t.co/9AP59aCPVM

Time to repurpose that old Apple Watch. https://t.co/NDZ3MsFM1L

Something must change in robotics! 😖 As more billion-dollar players enter the space, more systems are closing off. APIs move behind paywalls, platforms lock down, and robots turn into black boxes. @pollenrobotics is taking the opposite route. With Reachy 2, everything is open: the code, the CAD files, the ROS 2 interfaces, the communication stack. If you want to swap the camera, redesign the gripper, or fork the whole system, you can, and that’s the point. It creates a culture where universities can run real experiments without licensing drama, and startups can prototype without waiting for vendor permission. Clone it on GitHub. Add behaviors. Share datasets. Fork the models. Push improvements upstream. The @huggingface influence is real. It’s two open ecosystems aligning around the same idea: robotics should be built by more than a handful of companies. It was so cool to meet the team behind Reachy 2 in Bordeaux! 🔥 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

We’re excited that from a single video we can now create a sim-ready 3D scene—objects are interactable, separated, complete, and physically stable. We also show how these indoor digital twins enable VFX, robot sim, and even 3D printing!

ByteDance Seed releases GR-RL: Enabling robots to manipulate complex objects with human-like dexterity! • Tying shoelaces • Adjusting clothes • Moving soft objects How to achieve this? >Learning from human demonstrations, but only the essentials >Data augmentation for more flexible movements >Self-trial and error optimization for improved accuracy In the future, home assistant/elderly care robots will not only move things, but also do housework. Project page: https://t.co/OGQZe5jzdx ArXiv: https://t.co/xG273vNhhj

Holy s…it's game over. Which humanoid robot could beat the EngineAI T800? （Real footage, no CG, no AI, no speed-up.） https://t.co/T8O7xkkv4D

Amazon's FAR team has open-sourced a full-stack solution called Holosoma, providing an open-source, highly general framework for complex humanoid robot motion control and whole-body coordination. This can be seen as Amazon's capability stack prepared for future humanoid robots in the first mile of warehouse operations and the last mile of delivery (carrying and retrieving packages). Enabling humanoid robots to move and perform tasks with coordinated whole-body movements in reality remains a challenge. Especially in environments like Amazon warehouses and community delivery, for example, enabling humanoid robots to: • bend down to pick up items without falling • sit down and stand up • step onto stairs on one leg • walk on uneven ground This requires complex whole-body control, which is very difficult. Holosoma provides a unified reinforcement learning training framework specifically designed to address this problem. This framework is also applicable to humanoid robots of different sizes and other robot forms.

This seems Big Researchers have created a one dimensional “quantum wire” where mass and energy flow without friction or loss. Scientists at TU Wien have created a quantum system where transport has zero resistance, a "perfect flow" of mass and energy. By lining up thousands of ultracold rubidium atoms in a 1D atomic wire, they found that collisions don’t slow anything down. Instead, momentum is passed along like a quantum Newton’s cradle, allowing energy and mass to move indefinitely without fading. The experiment reveals a form of transport that defies normal thermodynamics and behaves like a perfect quantum conductor, it offers a new 'window' into how resistance can vanish at the fundamental level.

Video AGI has also been reached! Runway Gen-4.5 looks insane!🤯

DeepSeek V3.2(special) is a massive release! I think some people don't understand just how massive this release is! - They are the first, even ahead of OpenAI and Google, to release a Gold IMO 2025, CMO 2025, IOI 2025, and ICPC World Finals model! Everyone now has access to such an outstanding model. - The claim that open source is eight months behind closed source seems to be refuted. Open source is catching up with closed source and is only slightly behind now - and at the same time, it's increasing the pressure on US AI companies (I also recently read an article that more and more Silicon Valley companies are relying on open-source models). DeepSeek-V3.2 is an open-source reasoning and agentic model that closes much of the gap to frontier systems like GPT-5 and Gemini-3.0-Pro by combining a new sparse attention mechanism with aggressively scaled RL post-training. -DeepSeek-V3.2’s RL post-training budget exceeds 10% of pre-training compute, a key driver of its GPT-5-level reasoning performance (paper) -The agentic pipeline synthesizes over 1,800 environments and 85,000 complex prompts for tool-use RL, covering search, coding, and general agents (paper) This is nuts!

Deepseek casually drops a 685 Billion Parameter model on Hugging Face with MIT license! The higher variant performance on par with GPT-5 and Gemini 3.0 Pro! And yet, this model is open weights coming with a Technical Paper! https://t.co/hTKte8ENDj

Crafting the impossible! 🦌 An amazing display of engineering by this CNC machine. Look at this precision, design complexity, and technical mastery. In recent times in robotics, more and more complex parts are being created on CNC machines, and printed. Also, the degree of complexity will grow. CNC skills will be at a premium! The next time you want to craft a part on your machine and want to give up - remember this video! ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

This Stanford University paper just broke my brain. They just built an AI agent framework that evolves from zero data no human labels, no curated tasks, no demonstrations and it somehow gets better than every existing self-play method. It’s called Agent0: Unleashing Self-Evolving Agents from Zero Data via Tool-Integrated Reasoning And it’s insane what they pulled off. Every “self-improving” agent you’ve seen so far has the same fatal flaw: they can only generate tasks slightly harder than what they already know. So they plateau. Immediately. Agent0 breaks that ceiling. Here’s the twist: They spawn two agents from the same base LLM and make them compete. • Curriculum Agent - generates harder and harder tasks • Executor Agent - tries to solve them using reasoning + tools Whenever the executor gets better, the curriculum agent is forced to raise the difficulty. Whenever the tasks get harder, the executor is forced to evolve. This creates a closed-loop, self-reinforcing curriculum spiral and it all happens from scratch, no data, no humans, nothing. Just two agents pushing each other into higher intelligence. And then they add the cheat code: A full Python tool interpreter inside the loop. The executor learns to reason through problems with code. The curriculum agent learns to create tasks that require tool use. So both agents keep escalating. The results? → +18% gain in math reasoning → +24% gain in general reasoning → Beats R-Zero, SPIRAL, Absolute Zero, even frameworks using external proprietary APIs → All from zero data, just self-evolving cycles They even show the difficulty curve rising across iterations: tasks start as basic geometry and end at constraint satisfaction, combinatorics, logic puzzles, and multi-step tool-reliant problems. This is the closest thing we’ve seen to autonomous cognitive growth in LLMs. Agent0 isn’t just “better RL.” It’s a blueprint for agents that bootstrap their own intelligence. The agent era just got unlocked.

A 17-year-old just built a mind-controlled prosthetic arm for $300. Yes, $300. For something that usually costs $450,000. Let that hit you. A teenager, working from home, used AI, cheap materials, and 23,000 lines of code to build a device that reads brain signals without surgery, without implants, and without a $450K price tag. This is not a feel-good story. It’s a warning shot. How can a high school student build something 1,500× cheaper than the industry standard? What does that say about innovation? About pricing? About who gets access to life-changing technology? Of course, medical prosthetics are expensive for real reasons: materials, testing, regulation, customization. But let’s be honest — not all of that justifies a half-million-dollar price. This story exposes a simple truth: The future of accessibility won’t come from the system. It will come from the outsiders who dare to challenge it. If a 17-year-old can match top-tier prosthetics for a fraction of the cost… why aren’t these solutions available to the millions who need them? What do you think — breakthrough moment or the start of a bigger revolution? #AI #Innovation #Healthcare #Accessibility #FutureOfTech

Beijing just announced they will PAY foreign scientists to do whatever research they like in China!! Its for international science scholars that graduated from top 500 global universities. China will let you pick your own research topic and get 200k RMB of funding per year to do it, as well as a big salary and free housing too!! China letting you come here for big money, working with the best talent, to let you do whatever research you like. Completely different approach than in the West who don’t really have any programs like this for foreign talents. Places like the UK and US right now very much have uncertain quotas for visas, often far slower and more hostile-feeling than advertised like this. China's move here is different, the scheme literally says “foreign scholars only” as they want the best researchers here. Very smart strategy by China and one I think will see them overtake much of Europe and the US in terms of innovation and research. I think we need to wake up in the UK for sure on this!!!

Meanwhile in China 🇨🇳 where William & I spotted a ton of dual gunning trucks, the triple gun charging truck is now arriving to the market: first plug is exclusively for pack cooling according to @ChinaEV_Eng_Lif the ramp up in 35 seconds to nearly 2000A fast charging🤯 1.39MW https://t.co/FU8jtyYGFT

Imagine you no longer need any network or any radio frequency, just a single photon. Cisco has a working chip that generates 200 million quantum entanglements per second. In theory a single photon can encode the ENTIRE corpus of the world’s information. AI will be a photon. https://t.co/GGhscAoiA9

Just found this database of 91 humanoid robots. Filter by: - Price ($0–$200k) - Max speed (up to 22 km/h) - Payload strength (up to 70 kg) - Navigation performance - Manipulation performance Check out: @Humanoidguide https://t.co/yMcjdh1JXZ

Many people made me aware of Kimi agentic slides using Nano Banana Pro. Seems great! Will try that now.

I just open-sourced 100+ AI Agent and RAG tutorials. 100% free and Opensource (80k+ stars) with step-by-step tutorials. No excuses. Just Start building today. https://t.co/j3Rg2VA2t6

Mathematical superintelligence is coming, faster than you imagined

General purpose robots are coming fast A new framework (RGMP) lets humanoid robots understand object geometry, choose the right manipulation skill, and act with 5x more data efficiency. 87% success rate even on unseen tasks. https://t.co/6hV4uaJnjN

What if the most dangerous thing about AI isn’t that it kills us… but that it quietly kills capitalism itself — and we’re all too traumatised by the 20th century to even dare imagine what comes next without screaming “communism!”? Eric Weinstein just went nuclear in a 2:38 clip with Nicole Shanahan that will be studied in 2050: “We all know how to be terrified of AI. But the future is also going to be unbelievably cool.” His thesis: • AI is about to shift trillions from labor → capital at a speed never seen in history • Adam Smith, Marx, AND Hayek all become simultaneously obsolete • The real emergency isn’t saving capitalism or fearing communism • It’s that nobody is allowed to host a single conference titled “Capitalism Will Not Survive AI… and That Could Be the Best Thing That Ever Happened to Humanity” Nicole: “People are terrified to talk about life beyond capitalism because it instantly sounds Marxist.” Eric: “Exactly. We’ve lost the ability to invent brand-new economic space. We’re like pre-1989 wheeled luggage: everyone knows it always falls over, so nobody funds the guy who finally fixed the center of gravity.” We are living through an economic Berlin Wall moment and pretending the wall isn’t there. Watch these 2:38 minutes. Seriously. It’s the most important conversation almost nobody is having. Who else is ready to invent the economics that hasn’t been named yet? 2:38 video below

FindME: A CLI tool for searching social media and online profiles linked to a username. It’s ideal for reconnaissance, digital footprint verification, or checking username availability. GitHub: https://t.co/2lrkbdvNeU https://t.co/rEFtyyqT5p

I still can’t believe this is free. Most bootcamps are charging $3,000 to teach you outdated material. Meanwhile, @huggingface is giving away the state-of-the-art curriculum for $0. • Agents? ✅ • Robotics? ✅ • The new MCP standard? ✅ Check this. Bookmark.👇 https://t.co/HK5zpvxFmW

8 key skills to become a full-stack AI Engineer: (free/open-source resources below) https://t.co/HFi5ewwNpq

Anthropic seems to have found something that has allowed Claude Opus to develop significantly. Not only does it perform considerably better than Sonnet 4.5, but it's also two-thirds cheaper. Almost a contradiction - but Anthropic has achieved something significant internally.

This may be the biggest breakthrough of the year Chinese Physicists Just Solved Quantum Computing’s Hardest Challenge - Protecting Qubits From Environmental Noise. Pan Jianwei, often called "the father of quantum" and his team used their Zuchongzhi 2 quantum processor to create a 'quantum Lego block' that simply refuses to fall apart, even when shaken by noise or errors. They engineered an exotic new state of matter where quantum information gets locked into the corners of a system, protected by the deep math of topology, basically quantum armor. This is the first ever experimental realisation of non equilibrium higher order topological phases, a huge deal because it points to quantum bits that don’t break down, solving one of the biggest problems in quantum computing. And they did it on a 66 qubit programmable quantum chip, using just a 6x6 section to simulate and detect these ultra stable corner states, something that doesn’t naturally exist in nature. This work shows that today’s noisy quantum machines can already explore custom built quantum matter that pushes the frontier far beyond what classical physics can handle. Why it matters? If this scales, we’re talking practical, fault tolerant quantum computers that can power next level AI, drug discovery, and massive scientific simulations.

a 45-minute masterclass on creating amazing ui with gemini 3. asking a front dev to make something like this would cost a ton of fuss, cash, and sanity. now you can do it in a day. front is so cooked. https://t.co/lLpPnLKiqZ